acceptodds
Under review as a conference paper at ICLR 2027

Jigsaw-CRL: Recovering Global Latent Causal Order from Fragmented Multi-Client Interventions

Abstract

Causal representation learning (CRL) aims to recover latent causal variables and their structural relations from high-dimensional observations. Existing CRL methods typically assume that all environments are defined over the same latent variables, or at least share a common latent representation space. We study a fragmented multi-client setting, where multiple clients interact with the same global latent causal system but each client only accesses and intervenes on a subset of the latent variables. In this regime, marginalizing unused latent variables induces bidirected edges, so a single client no longer admits a node-wise latent causal graph, and the global latent causal order must be recovered by assembling client-specific structural fragments. We propose **Jigsaw-CRL**, a framework for recovering global latent causal order from such fragmented interventions. Under soft interventions, differences between precision matrices across environments exhibit a low-rank structure governed by latent ancestor relations. This enables recovery, for each client, of a block partition, the corresponding block-level ancestral order, and latent subspaces, and then assembly of these fragments into the global node-level latent causal order. We establish identifiability guarantees, develop practical algorithms, and validate the framework on synthetic data. Our codes are available on https://anonymous.4open.science/r/code-for-Jigsaw-CRL-7B26

Primary area:
causal reasoning
Keywords:
causal representation learning, multi-client learning, causal discovery

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_EC6WOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates structure identifiability in causal representation learning. The authors consider a multi-domain setting, in which in each domain observational and interventional data is available. The data is generated by linear functions of a subset of the latent variables of interest. The goal is to identify the causal relationships among the latent variables within each domain and then combine these relationships to recover a global causal graph over all latent variables. Under strong assumptions, the authors establish a sufficient condition for identifiability. Their proof is constructive and also leads to a finite-sample estimation method.

Critical strengths

The exact setup is to my knowledge not yet studied in the literature. In particular, allowing for bidirected edges between the latent variables only allows to recover parts of the latent structure, which seems interesting to me.

Critical weaknesses

To me, some key concepts in the manuscript are insufficiently discussed and lack clarity. It remains unclear where a new contribution is and which parts are only minor rewritten extensions of the existing literature. In particular, I have the following concerns:

  1. The main theoretical tool of the paper is the block RQ decomposition. This appears very similar to the partial order RQ decomposition developed in the paper "Linear Causal Disentanglement via Interventions" by Squires et al. (2023). Indeed, Definition 3 defines a block-level partial order which is the main input to Definition 4. The connection should be clarified. Moreover, the remaining proof techniques also appear to be very similar. This should at least be acknowledged.

  2. I have difficulties to follow the reasoning in lines 186 to 188. In particular, why are nodes not the right abstraction level for identification any more? To get intuition, it would be helpful to show in examples why this is the case and why recovering the proposed blocks in Definition 4 is the correct level. That is, why can we not recover relations within the proposed blocks? And what is the intuition on why the blocks are exactly the right object? I would also appreciate examples on how the blocks look like.

  3. In Assumption 1 the authors require that "interventions from different clients that target the same global latent node share the same label". I have several comments on this: First, while I can imagine what it should exactly mean, it should be formalized. Second, I think that this is a very strong assumption. In particular, we need to know a priori which interventional data sets corresponding to different clients are obtained from interventions on the same node. This assumption should be properly discussed. What are real-world examples where it could be defended, and where are the limitations, that is, in which scenarios does it not hold? To what extent is the assumption necessary? Third, together with Assumption 3 and the conclusion from Theorem 1, Theorem 2 then appears to be trivial to me. Since coming up with sufficient conditions for a desired results is in some sense always possible, it should be better discussed why these assumptions are necessary.

  4. I am slightly concerned about the simulation study in the main part since this is only for three specific sparse graphs with fixed parameters and fixed Ok′sO_k's, and the soft interventions have much higher magnitude than the original parameters. Simulations as in Section N.6 appear more interesting to me, although the description is missing details such as the strength of the interventions. The case study in Section N.7 misses references for the data, the underlying ground true latent graph, and an exact description of the available data. I did not find code for Appendix N.6 and N.7 in the provided repository.

  5. Some minor weaknesses: While it is of course okay to study special cases, it should be made clear in the introduction that all relations are assumed to be linear. Moreover, it should be emphasized somewhere that the latent graph is assumed to be acyclic and that it is assumed to be known which one is the observational regime. Also, please give a definition for Hasse edges.

Questions for authors
  1. Is the block RQ decomposition a special case of the partial order RQ decomposition by Squires et al. (2023)? If not, why not?

  2. As in your setup, consider one client that only has access to a subset of the latent variables but there is interventional data for each of the latent variables. How is this related to assuming that the client has access to all latent variables but only to interventional data obtained from interventions on a subset of the latent variables?

  3. Isn't there some scaling ambiguity in the matrix HH? In the paper by Chandler et al. they rescale the largest absolute value in each row of HH to 11.

  4. In Theorem 1, why do we recover QQ and not HH? With this we still have the ambiguity of RR? Rather, I have the feeling that we could also identify RR up to permuting the columns by Pσ,kP_{\sigma, k} as in Theorem 3 in Squires et al.?

2 comments

As the discussion period approaches its deadline, we would sincerely appreciate hearing your thoughts on our response. We hope that the additional clarifications have addressed your concerns, and we welcome any further questions.

We also invite you to read our discussion with Reviewer 1CbJ, which further addresses the problem setting and the novelty of our model and theoretical results. We hope this discussion helps resolve your remaining concerns. If so, we would be grateful if you would consider revising your score upward.

Thank you again for your time and consideration.

I thank the authors for their response. The example in Section 2 is helpful and substantially improves the exposition. I will therefore raise my score by one point. Concrete examples and explanations of the underlying ideas are essential to making theoretical contributions accessible, and the added example is a welcome step in this direction.

Reviewer_5eZUOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies causal representation learning from fragmented multi-client interventions, where each client observes a linear mixture of only a subset of shared latent variables. It proposes a block-based characterization of locally recoverable structure and subspaces, algorithms using low-rank precision perturbations, and a procedure for assembling local information into a global causal order. Recovery guarantees are stated under intervention genericity and coverage assumptions. Validation includes synthetic experiments with Gaussian and Laplace noise, sensitivity and scalability analyses, and a semiconductor-manufacturing case study.

Critical strengths
  1. The paper addresses a meaningful extension of causal representation learning. Studying clients that observe different mixtures of overlapping latent subsets captures an important limitation of shared-latent-space formulations. Understanding what remains recoverable under such fragmented access is a worthwhile research direction.

  2. The local-to-global perspective is conceptually valuable. The proposed framework distinguishes local ambiguity induced by marginalization from the global causal information available through complementary clients. Making this distinction explicit provides a useful foundation for studying causal recovery from fragmented data.

Critical weaknesses
  1. The block-level recovery guarantee has an unresolved gap. Definition 3 defines block ancestry through the existence of a directed ancestral pair across blocks, but this relation need not be transitive: consecutive relations may involve different nodes within the intermediate block. Algorithm 2 nevertheless propagates ancestry through intermediate blocks, potentially adding relations absent from the defined target. This directly affects Theorem 1 and requires reconciliation between the definition, algorithm, and proof.

  2. The experiments lack comparative evaluation. The paper reports performance across sample sizes and settings without comparing against existing methods or simple alternatives. For example, applying a single-client method separately and naively combining its outputs would help quantify the benefit of the proposed block-aware approach, with its assumption mismatch explicitly acknowledged. Without such comparisons or component ablations, the experiments provide limited evidence of the practical gains from the proposed methodology.

Questions for authors
  1. How is the ancestor-propagation step in Algorithm 2 justified under Definition 3, whose existential definition of block ancestry need not be transitive?

  2. Are there comparative or ablation results showing the benefit of the proposed block-aware recovery and assembly over simpler alternatives?

3 comments

As the discussion period approaches its deadline, we would sincerely appreciate hearing your thoughts on our response. We hope that the additional clarifications have addressed your concerns, and we welcome any further questions.

We also invite you to read our discussion with Reviewer 1CbJ, which further addresses the problem setting and the novelty of our model and theoretical results. We hope this discussion helps resolve your remaining concerns. If so, we would be grateful if you would consider revising your score upward.

Thank you again for your time and consideration.

Thank you for the thoughtful revisions and the additional baseline comparison. The corrected block-order definition addresses my main theoretical concern, and I also appreciate the non-cancellation assumption and ancestor-closure correction (reply to Reviewer 1CbJ). The new comparison provides useful evidence for the proposed approach, although comparative performance with finite samples remains insufficiently explored. On balance, I am updating my rating to weak accept. Thank you for the careful responses and efforts to improve the paper.

Thank you for your response. We are deeply grateful for the time and effort you have devoted to reviewing our work. Your suggestions have been very helpful in further improving the quality of our manuscript.

Reviewer_1CbJOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3(+1)
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates linear transformation and linear SEM CRL setting under marginalized/partially observable latent causal models. The contribution is separated in two stages: What can one say given one such view ("client"), and how can we aggregate multiple views into a coherent model. The single-view analysis follows Squires et al. 2023 except under the marginalized model parameters (stated in Proposition 1). The authors show that the induced graph's cycles ("blocks") can be recovered up to their ancestors, together with the ancestral graph among them, using the usual projected-subspace arguments of Squires et al. 2023 or e.g. Varici et al. (2024 Neurips; "Linear causal representation learning from unknown multi-node interventions"). The aggregation step depends on an extremely strong assumption on data availability, and establishes that elementwise ancestral relations -- but NOT the variable values, even up to ancestors -- can be recovered.

Critical strengths

The per-client perspective is novel in comparison to existing work: This is, in essence, the confounded latents problem, which is relatively understudied. The resulting identifiability classes are intuitively reasonable --- cycles are often hard to resolve (though not intrinsically impossible, see Morioka & Hyvärinen, 2024) --- and as far as I know novel i.e. not equivalent to any existing result. Approach is constructive.

Critical weaknesses

Global aggregation is order only and requires a quite strong grouping/observability requirement. Even then, global recovery result does not refine/improve the latent variable estimates, even though they are able to refine the graph estimate.

This means the paper's merit should be mostly evaluated on what it does in the per-client case. The per-client algorithm and proofs are mostly correct. For example, transition from line 1322 in Step 2 to the argument in Step 3 is not correct (it could be smaller than full ancestors), but it is salvagable by minor changes e.g. applying transitive closure to A\mathcal{A} before proceeding. Similarly, there are a couple of places where rank-is-nonzero-unless-finetuned argument is incomplete (i.e. actually requires a new assumption) e.g. around 1874, we can fine-tune the model weights s.t. observational AkA_k entry can be zero despite nominally having a induced dependency thus nonzero dkd_k. Since I have found a couple errors myself, I would recommend a thorough vetting of all proofs in revision/before resubmission.

However, the tools are not too novel: the precision matrix approach -- as well as the matrix decomposition idea -- for linear+linear CRL setting follows almost verbatim after Proposition 1 from Squires et al. 2023. Similarly, the subspace-projection-based analysis is pretty standard in literature (see above for a couple references). I think the problem definition -- modulo the issue of not having a global recovery/aggregation setup for variables -- is interesting, but I am not sure if the current amount of novelty is enough.

Questions for authors
  1. Is it possible at all to get a global variables recovery result as well?
  2. Can you please fix all proof issues in the paper?
  3. Given that the results most closely resemble the latent confounding setting, can the authors provide a more complete literature review and list their differences/improvements on it?

I am on the fence on whether to recommend acceptance given the minor lack of technical and theoretical novelty, even if the authors could convince me of the correctness of their updated proofs. I am willing to increase my score if they manage to, and both weaknesses are strengthened.


Post-diacussion, I have been convinced about the relative novelty of the setting, both w.r.t. its setting and its recovery objectives. Thus I am increasing my informativeness and overall scores from 2 to 3.

9 comments

Answer to Q1

Under our current assumptions, we can use the client-level results to recover the global causal order. At the level of latent variables, our guarantee is client-specific: each client recovers its latent variables up to mixing within blocks and with ancestor blocks.

Regarding the grouping assumptions, some cross-client correspondence is necessary for global recovery, as discussed in Appendix C and in our response to W3 from Reviewer EC6W. In particular, without the aligned intervention labels in Assumption 1, the same two unlabeled client-level fragments can be consistent with either the global graph 1→2→31 \to 2 \to 3 or the global graph 1←2→31 \leftarrow 2 \to 3. These graphs have different ancestral orders even up to relabeling. Assumption 3 addresses a separate issue, namely whether the clients collectively cover the relations needed for global recovery. We show that this condition holds with high probability under the Erdős–Rényi graph model considered in the paper, and the simulations in Appendix N further illustrate this result.

The scope of our graph recovery guarantee also reflects our choice of soft interventions, as discussed in our response to Q4 from Reviewer EC6W. Even for a single client with access to all latent nodes, Squires et al. (2023, Appendix J) give counterexamples showing that soft interventions can leave both the unmixing matrix HH and the exact DAG unidentified. Intuitively, the data identify the factor QQ in the decomposition H=RQH=RQ, while the remaining ambiguity in RR allows a latent variable to mix with its ancestors. Distinct DAGs can therefore induce the same observed distributions while sharing the same ancestral order. In this sense, our global order result reaches a worst-case limit of the soft-intervention setting. Hard interventions provide a natural direction for studying exact global recovery in future work. So in the current version with soft intervention, we can not directly obtain the exact global causal structure. But this is the limitation of soft intervention, not the limitation of our algorithm.

Our main contribution concerns the bidirected edges and block structure that arise when different clients observe different subsets of the latent variables. As explained in our response to W2 from Reviewer EC6W, the shared linear setting makes some of the analytical tools of Squires et al. (2023) applicable here. The new difficulty is that their node-wise recovery criterion can fail even after all ancestor blocks have been accounted for. Our analysis therefore has to show why the nodes in a block can be recovered jointly and why no nonempty proper subset satisfies the same recovery criterion. This is nontrivial analysis.

Answer to Q2

Thank you for checking the proofs so carefully. Your comments helped us identify an omitted closure step in Algorithm 1 and a cancellation under marginalization that our original assumptions did not explicitly exclude. We address each point below.

1. The ancestor set at line 1322. As you correctly observed, after the pruning loop in Algorithm 1, A\mathcal A contains all block-level parents of the current block CtC_t, but it may contain only some of its more distant ancestors. This is sufficient for the order update in line 14 of Algorithm 2, which records both the blocks in A\mathcal A and their previously recovered ancestors as ancestors of CtC_t. Since every ancestor of CtC_t is either a parent or an ancestor of a parent, this update recovers the full set of block ancestors.

However, Step 3 of the proof also uses the full ancestor set to construct the projection space WW. The original pseudocode omitted the closure step before this projection. We have therefore added the following assignment at line 9 of Algorithm 1, immediately before constructing WW:

A←the ancestor closure of A in I\mathcal A \gets \text{the ancestor closure of }\mathcal A\text{ in }\mathcal I.

We have revised Steps 2 and 3 of the proof accordingly. In particular, the inclusion at line 1322 of the original manuscript describes A\mathcal A immediately after pruning. After the new closure step, A\mathcal A contains exactly the block ancestors of CtC_t, as required in Step 3.

2. Cancellation near line 1874. We agree that the original assumptions did not exclude the case you identified. Assumption 3 in the revised manuscript (Assumption 2 in the original submission) addresses degeneracies caused by the intervention parameters. In your example, the cancellation occurs earlier: distinct path contributions cancel in the observational coefficient [Ak]ρk(t),ρk(v)[A_k]_{\rho_k(t),\rho_k(v)} when variables are marginalized out, even though the corresponding intervention perturbation is nonzero. To cover this case and its analogue for bidirected edges, we added the following assumption in Section 2.

Assumption 1 (No cancellation under marginalization). For each client kk and any distinct v,t∈Okv,t\in O_k, we assume that [Ak]ρk(t),ρk(v)≠0[A_k]{\rho_k(t),\rho_k(v)} \neq 0 whenever the global graph contains a directed path from vv to tt whose internal vertices all belong to MkM_k (including the case of a direct edge). We also assume that [Σk]ρk(v),ρk(t)≠0[\Sigma_k]_{\rho_k(v),\rho_k(t)}\neq 0 whenever some h∈Mkh\in M_k has directed paths to both vv and tt whose internal vertices all belong to MkM_k.

Like Assumption 3 in the revised manuscript (Assumption 2 in the original submission), we view Assumption 1 as a mild genericity condition. It rules out finely tuned parameter configurations that lead to exact cancellation under marginalization. In this sense, the excluded configurations can be viewed as a measure-zero subset of the parameter space. Assumption 1 therefore excludes precisely the type of fine-tuned cancellation you identified, while Assumption 3 addresses degeneracies in the intervention perturbations.

We thank you for bringing these points to our attention. We believe these revisions are minor, comparable in scope to correcting typos, and preserve the main ideas underlying our algorithm and proof. After making these revisions, we rechecked the proofs in full and found no further issues of this kind.

Answer to Q3

Thank you for suggesting a fuller comparison with the literature on latent confounding. Appendix A discusses most of the works cited below. For clarity, we have reorganized the comparison around two questions: which causal variables are directly observed, and what structure must be learned from the data? If the paper is accepted, we will incorporate the most relevant points from both parts into Appendix A.

1. Causal discovery under latent confounding. A substantial literature studies which causal relations among measured variables remain identifiable in the presence of unobserved common causes [1–3]. COmbINE [4] is particularly relevant because it also combines interventional datasets covering overlapping sets of variables. In that setting, the variables measured in each dataset are known. In our setting, the causal structure of interest is defined over latent variables, and each client observes an unknown mixture of only a subset of them. Xie et al. [5] also study latent causal structure, focusing on the conditions under which it can be identified from observational data alone in linear non-Gaussian hierarchical models. Their work falls within latent causal structure discovery, while our CRL setting uses soft interventions to identify both latent subspaces and their ancestral relations.

2. Causal representation learning with confounding or partial observability. Most intervention-based CRL methods in the single-client setting model a common set of latent variables by a DAG and typically exclude bidirected edges among those variables [6–9]. Three works are especially relevant to our setting.

  • Li et al. [10] allow bidirected edges among latent variables and take the latent causal graph as input. Given this graph and information about interventions, they characterize which latent variables can be disentangled. In our setting, the client-specific graph is unknown, and both the block partition and the ancestral relations between blocks must be learned from data.
  • Sturma et al. [11] learn a shared latent causal graph from unpaired domains. Their shared-graph identifiability result excludes edges between shared and domain-specific latent variables, as well as edges between latent variables specific to different domains. These restrictions rule out the source of confounding central to our setting: a latent variable omitted by one client can be a common cause of two latent variables retained by that client.
  • Yao et al. [12] allow each view to contain a different subset of causally related latent variables. Their result identifies the latent information shared across simultaneously observed views, up to a smooth bijection. Their identification target is shared content, whereas ours is the block structure and ancestral order induced by each client's partial access to the causal system.

Our contribution is to characterize what causal structure remains identifiable when marginalization induces bidirected dependencies among latent variables observed only through client-specific mixtures. At the client level, we show that single-node soft interventions identify the block partition, the associated latent subspaces, and the ancestral order between blocks. Across clients, an ancestral relation hidden within a block for one client may be resolved by another. We give conditions on intervention-label alignment and client coverage under which these complementary relations identify the global ancestral order among individual latent variables.

References

[1] Richardson, T. S., & Spirtes, P. (2002). Ancestral graph Markov models. The Annals of Statistics, 30(4), 962–1030.

[2] Zhang, J. (2008). On the completeness of orientation rules for causal discovery in the presence of latent confounders and selection bias. Artificial Intelligence, 172(16–17), 1873–1896.

[3] Jaber, A., Kocaoglu, M., Shanmugam, K., & Bareinboim, E. (2020). Causal discovery from soft interventions with unknown targets: Characterization and learning. Advances in Neural Information Processing Systems, 33.

[4] Triantafillou, S., & Tsamardinos, I. (2015). Constraint-based causal discovery from multiple interventions over overlapping variable sets. Journal of Machine Learning Research, 16(66), 2147–2205.

[5] Xie, F., Huang, B., Chen, Z., Cai, R., Glymour, C., Geng, Z., & Zhang, K. (2024). Generalized independent noise condition for estimating causal structure with latent variables. Journal of Machine Learning Research, 25(191), 1–61.

[6] Squires, C., Seigal, A., Bhate, S. S., & Uhler, C. (2023). Linear causal disentanglement via interventions. Proceedings of the 40th International Conference on Machine Learning, 202, 32540–32560.

[7] Buchholz, S., Rajendran, G., Rosenfeld, E., Aragam, B., Schölkopf, B., & Ravikumar, P. K. (2023). Learning linear causal representations from interventions under general nonlinear mixing. Advances in Neural Information Processing Systems, 36.

[8] von Kügelgen, J., Besserve, M., Wendong, L., Gresele, L., Kekić, A., Bareinboim, E., Blei, D. M., & Schölkopf, B. (2023). Nonparametric identifiability of causal representations from unknown interventions. Advances in Neural Information Processing Systems, 36.

[9] Varıcı, B., Acartürk, E., Shanmugam, K., Kumar, A., & Tajer, A. (2025). Score-based causal representation learning: Linear and general transformations. Journal of Machine Learning Research, 26(112), 1–90.

[10] Li, A., Pan, Y., & Bareinboim, E. (2024). Disentangled representation learning in non-Markovian causal systems. Advances in Neural Information Processing Systems, 37.

[11] Sturma, N., Squires, C., Drton, M., & Uhler, C. (2023). Unpaired multi-domain causal representation learning. Advances in Neural Information Processing Systems, 36.

[12] Yao, D., Xu, D., Lachapelle, S., Magliacane, S., Taslakian, P., Martius, G., von Kügelgen, J., & Locatello, F. (2024). Multi-view causal representation learning with partial observability. International Conference on Learning Representations.

Comments on the paper

3 comments

I thank the authors for the efforts to address my comments. However, I remain concerned that the manuscript would still require substantial improvements before it would be suitable as a camera-ready paper. In particular, it does not meet my bar on clarity and presentation.

Moreover, not all of my concerns are adequately addressed:

W1: I am still concerned why the work by Chandler et al was not clarified and acknowledged in the initial version. W2: The intuition that blocks are strongly connected components in the auxiliary graph is helpful and should be explained in the paper. However, my questions should still be explained better. That is, why can we not recover relations within the proposed blocks? And what is the intuition on why the blocks are exactly the right object? W3: I would appreciate such examples in the main paper. W4: I can not tell whether my concerns will be appropriately addressed in a revised version. W5: But if “Hasse edge” is defined in Assumption 3, why do you use the (not well-known) term already in line 85?

Thank you for your helpful suggestions. We have revised the manuscript in response to your follow-up comments and explain the changes point by point below. We hope these revisions address your concerns, and we welcome any further questions.

Answer to W1

In the original manuscript, we cited Squires et al. (2023) in the Introduction and in the discussion accompanying Assumption 2. We also cited their work in the literature review, which appears in Appendix A due to space constraints. In the revised manuscript, we cite their work again in a new paragraph at the end of Section 2, Explanation and example of the block structure. There, we explain how their approach uses the topological order of a DAG, why bidirected edges prevent the same node-wise identification strategy from applying directly, and why extending recovery to blocks is nontrivial.

Answer to W5

Thank you for pointing this out. We apologize for leaving the term undefined at its first occurrence. In the revised manuscript, we give an intuitive definition when the term first appears on line 85 and refer the reader to Assumption 3 for the formal definition. The revised text reads:

a Hasse edge (an edge whose endpoints have no other node strictly between them in the causal order, as defined in Assumption 3)

Answer to W2–W4

We added the paragraph Explanation and example of the block structure at the end of Section 2 to clarify how blocks are defined, why their internal order cannot be recovered node by node using the layer-wise criterion, and in what sense a block is the smallest identifiable unit. For convenience, we give the explanation here.

Consider the graph G1\mathcal{G}_1 with directed edges 1→2→3→4→51 \to 2 \to 3 \to 4 \to 5, and the graph G2\mathcal{G}_2 obtained by adding the bidirected edge 2↔42 \leftrightarrow 4 to G1\mathcal{G}_1. In the DAG setting considered in much of the existing literature, nodes can be identified layer by layer in topological order under suitable interventions (Squires et al., 2023; Varıcı et al., 2025). The intuition is that once the directions associated with a node’s parents have been projected out, an intervention on that node can be distinguished from interventions on nodes with unresolved parents. For example, in Squires et al. (2023), the projected difference between the interventional and observational precision matrices has rank one when all parents of the target have been accounted for. In G1\mathcal{G}_1, this criterion first identifies node 11, which has no parents. Projecting out its direction then reveals node 22, the root of the remaining graph. Repeating this step recovers the entire order.

In G2\mathcal{G}_2, the bidirected edge disrupts this layer-wise procedure by linking nodes in different layers through an off-diagonal entry of the noise covariance matrix. We may still identify node 11, but projecting out its direction no longer suffices to isolate node 22. Through the bidirected edge, information from node 44 and its ancestors remains relevant to identifying node 22. This dependence works in both directions: information from node 22 and its parents also affects the identification of node 44. The node-wise strategy therefore breaks down because these dependencies cannot be removed one layer at a time.

Reachability in Definition 1 describes which nodes carry information relevant to identifying a given node. When nodes are mutually reachable, the dependencies induced by bidirected edges and directed ancestry form a cycle in the identification procedure. This motivates the blocks in Definition 2. Specifically, in G2\mathcal{G}_2, isolating node 22 would require accounting for nodes {1,3,4}\{1,3,4\}, isolating node 33 would require accounting for {1,2,4}\{1,2,4\}, and isolating node 44 would require accounting for {1,2,3}\{1,2,3\}. Thus, even after node 11 has been accounted for, none of the three remaining nodes in {2,3,4}\{2,3,4\} can be identified first. Any nonempty proper subset of {2,3,4}\{2,3,4\} still has an unresolved dependency on another member of the block. In this sense, the block is the smallest unit that can be identified at this step.

If we consider {2,3,4}\{2,3,4\} as a whole, its only external ancestor is node 11. Section 3 proves that, under our assumptions, projecting out the direction of node 11 allows us to identify {2,3,4}\{2,3,4\} jointly. After projecting out the directions associated with both node 11 and this block, we can identify node 55. Establishing block-level recovery requires showing both that the full block can be isolated and that no proper subset satisfies the same recovery criterion. This is what makes the extension beyond node-wise recovery nontrivial, as detailed in our proofs.

The block structure also has a graphical interpretation. Construct an auxiliary graph by replacing each bidirected edge with two oppositely directed edges. Reachability in Definition 1 then corresponds to ancestry in the auxiliary graph, and the blocks in Definition 2 are exactly its strongly connected components. In G2\mathcal{G}_2, for example, the auxiliary edge 4→24 \to 2 closes the cycle 2→3→4→22 \to 3 \to 4 \to 2. This explains why node 33 belongs to the same block as nodes 22 and 44, even though it has no incident bidirected edge. The two auxiliary directions capture the fact that the dependence represented by a bidirected edge can be traced in either direction.

References

  1. Squires, C., Seigal, A., Bhate, S. S., & Uhler, C. (2023). Linear Causal Disentanglement via Interventions. Proceedings of the 40th International Conference on Machine Learning, 202, 32540–32560.

  2. Varıcı, B., Acartürk, E., Shanmugam, K., Kumar, A., & Tajer, A. (2025). Score-based Causal Representation Learning: Linear and General Transformations. Journal of Machine Learning Research, 26(112), 1–90.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.