acceptodds
Under review as a conference paper at ICLR 2027

When Causal Structure Replaces Replication: Reference-Free Federated Causal Auditing

Abstract

In federated causal analysis, different data holders may answer different fixed conditional-independence (CI) questions without a common reference answer, so the same reports can support different explanations of the true CI answers and source alignments. We ask when relationships among CI statements can resolve this ambiguity without additional cross-source answers and when such answers remain necessary. For any specified set of possible CI answer patterns, we exactly characterize the surviving ambiguities, when relative source alignment is identified, and, under any fixed set of allowed cross-source requests, the minimum number of additional answers needed, including when identification is impossible. Holding the queries, sources, reporting graph, reporting mechanism, and allowed requests fixed, we prove that a particular four-variable pattern of CI changes and its complement cannot be the exact difference between two DAG-consistent causal explanations, even inside a larger DAG, while the same change is realizable by regular Gaussians. Repeating the construction across disjoint four-variable blocks gives a linear separation with only a linear number of fixed CI questions: DAG structure needs no additional cross-source answers, whereas regular Gaussians need exactly one per complete block, even when every missing source–query answer may be requested. Controlled exact comparisons over the complete ambient-DAG quartet family show that the replication advantage is common across broader reporting patterns, while a separate repeated-task analysis shows an analogous identification benefit under asymmetric stochastic source errors.

Primary area:
causal reasoning
Keywords:
causal inference, conditional independence, graphical models, identifiability, source alignment, cross-source replication

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_pJMMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the problem of federated causal auditing without reference labels. It leverages conditional independence (CI) structural constraints to identify alignment ambiguities across data sources and derives the minimum cost of redundant queries required when operating across multiple sources.

Critical strengths
  1. Establishes a theoretical link between the identifiability of source alignments and the minimum cost of redundant queries.
  2. Demonstrates the advantages of using directed acyclic graph (DAG) structures—compared to general Gaussian models—in reducing redundant queries.
Critical weaknesses
  1. The method assumes that different data sources share the same underlying true CI pattern, which limits its applicability to scenarios involving data heterogeneity.
  2. Experiments are primarily based on synthetically constructed CI patterns, lacking validation using standard causal networks or real-world datasets.
  3. The paper lacks an analysis of computational complexity and scalability for large-scale scenarios.
Questions for authors

See above.

3 comments

Thank you for your feedback. We’ll shortly share a revised manuscript that clarifies the shared CI truth assumption and expands the discussion of benchmark validation and computational complexity, and we’d be happy to clarify any remaining questions. If you find that the rebuttal and revisions address your concerns, we’d be grateful if you would consider raising your rating.

Thank you again for your thoughtful comments. We have uploaded a revised manuscript that clarifies the shared CI truth assumption and its limitations under data heterogeneity, adds ALARM oracle and Sachs measured-data evaluations, and expands the computational analysis with exact-versus-greedy comparisons and factorised-family scalability experiments.

We would be grateful for your feedback on these additions and clarifications. If you have any remaining questions about the theoretical analysis or experimental results, we would be happy to discuss them.

AuthorsOpenReview ↗

A Sincere Follow-up Before the Discussion Deadline

Dear Reviewer pJMM,

Thank you again for your time and thoughtful review. With approximately one day remaining in the discussion period, we sincerely hope to hear your thoughts on our responses and revised manuscript. Your feedback is very important to us, and we would be grateful for any further comments or remaining concerns you may wish to share.

Thank you very much for your time and consideration.

Reviewer_dWz3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper addresses the important problem of causal discovery with the use of conditional independence tests coming from different sources. The authors find the minimum number of required tests to resolve ambiguities between the sources and provide rigorous analysis. They also provide results on the special case of four blocks. Overall, a strong paper.

Critical strengths
  • The problem of finding a causal structure by the use of conditional independence tests coming from different sources (federated) is interesting and important. Finding the minimum number of required tests in a way to resolve the conflicts between the sources is a great step to solve this problem, which the authors took.
  • The paper has several rigorous results.
Critical weaknesses
  • Presentation is not smooth enough; it's targeting an expert in the field of causality + federated learning. This is the greatest concern. I had a hard time understanding the text starting from Page 3. The authors are encouraged to provide a visual example where the concepts in page 3 and after are illustrated. They can then revisit the example to illustrate Ambiguity geometry and Proposition and/or Theorem 1. I am willing to increase my score if such an example is provided.
  • This is not a general result, but just a starting point as it's limited to just four blocks: "Complete-action linear evidence separation." I wouldn't frame it any greater than what it is; although the word "evidence" here helps.
  • Some of the potentially related federated and data distributed work in the following review are missing: Ramazi, P., Fast, A., Fernandopulle, C., Haratian, A., Sanjaroon, V., & Alkema, A. (2026). Structure Learning in Bayesian Networks: A Systematic Pedagogical Survey.
Questions for authors
  • "gold labels" I suppose it means true labels?
  • difficult to sparse: "In the hospital example, one disconnected group can be paired with a compatible change in the underlying CI truths without changing any reported answer"
1 comment

Thank you for suggesting a visual example. We’ll shortly share a revised manuscript with a worked example and a clearer explanation of the scope of the linear separation result, and we’d be happy to clarify any remaining questions. If you find that the rebuttal and revisions address your concerns, we’d appreciate your considering a higher rating.

Reviewer_LrGqOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

In this paper, the authors addressed federated causal analysis where conditional independence (CI) statements are obtained in different distributed sources and the ambiguity among the statements need to be resolved. The authors show their theoretical claims for arbitrary DAGs and demonstrate empirical performance for different setups.

Critical strengths

The authors address an important problem: Federated causal discovery. I appreciate that the authors targeted the problem from a theoretical perspective. They provided multiple theoretical claims and proofs. I appreciate their experimental analysis which can be improved, however.

Critical weaknesses

Below I share my concerns:

  1. Many part of the paper including the abstract and introduction is really hard to understand as described in a convoluted way. The authors should consider writing it in a simple manner with proper examples.
  2. The addressed problem is not described precisely and not clear.
  3. In the introduction, some terms are unclear and mentioned without proper example or discussion. For example: ambient DAGs, source–query observation graph, dimension-minimal four-variable causal exclusion, disjoint four-variable blocks, etc.
  4. The authors should discuss existing federated causal discovery approaches such as [1,2]
  5. Markov and faithfulness property in Assumption 1 is not defined properly.
  6. The experiment setup is very limited. Also, the approach was not evaluated on any real-world dataset. Networks such as ALARM and Sachs are mentioned in the appendix but not evaluated on. Also, more baselines are required.

References:

  1. Mian, Osman, et al. "Nothing but regrets—privacy-preserving federated causal discovery." International Conference on Artificial Intelligence and Statistics. PMLR, 2023.\
  2. Yang, Dezhi, et al. "Federated causality learning with explainable adaptive optimization." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 38. No. 15. 2024.
Questions for authors

Questions:

  1. What do the authors refer to by “disjoint four-variable blocks”?
  2. What do the authors refer to by ambient causal system or ambient dag? Did the authors provide specific definition of such a causal system in the paper?
  3. What does happen if Assumption 1 fail?
  4. Can the authors cite some existing works in federated causal discovery literature from which they borrowed their machinery? i.e. There should be some work on top of which the proposed method is built.
  5. How did the authors calculate the values in eq 47?
  6. How does the proposed approach perform on ALARM and Sachs?
4 comments

Thank you for your detailed feedback. We’ll shortly share a revised manuscript with a clearer problem statement and a fuller discussion of the experiments, and we’d be happy to clarify any remaining questions. If you find that the rebuttal and revisions address your concerns, we’d be grateful if you would consider raising your rating.

Thank you again for your detailed feedback. We have uploaded a revised manuscript with a clearer problem formulation, more accessible explanations of the terminology and CI assumptions, and expanded experimental evidence, including ALARM oracle evaluations, Sachs measured-data results, and exact-versus-greedy comparisons. We have also clarified the mathematical foundations and related work in response to your questions.

These revisions incorporate the clarifications and additional experiments discussed in our rebuttal. We would appreciate your thoughts on the updated manuscript and would be happy to clarify any remaining questions.

I thank the author for their effort to address my concerns. However, I am still concerned about the language of the paper and not sure how interpretable it will be to the readers. Although the paper enlisted "causal reasoning" as its primary area, it does not use standard causal terminologies. From very concise discussion of different terms such as faithfulness and Markov property, it seems terms might have been used without proper understanding and exploration. I would request the authors to use already established concepts in causality to improve the interpretability of their paper.

For example, in the rebuttal, the author mentioned that "For example, an unqueried H on A<-H->C changes the relationship between A and C." Do the authors refer to unobserved confounders here? Also, I would request the author to share an existing work that uses "Ambient-DAG".

For the above reasons, I would keep my score unchanged. I wish good luck to the authors.

AuthorsOpenReview ↗

Clarification of Causal Terminology and Assumptions

Thank you for your follow-up. The paper uses standard causal DAGs, d-separation, the global Markov property, and faithfulness in their conventional meanings. Please see these passages in the revised manuscript:

  • Section 3, immediately after Assumption 1: “The Markov property states that d-separation in the ambient DAG implies statistical CI in its population distribution; faithfulness gives the converse.”
  • Section 4.3, before Theorem 2: the quartet family comprises “all complete Q₄ profiles obtained by restricting the d-separation model of an arbitrary finite ambient DAG to the quartet U.”

“Ambient DAG” is a descriptive label for the full underlying DAG. Sections 4.3 and Appendix D.1 cite Richardson and Spirtes (2002), Ancestral Graph Markov Models, for the underlying marginalisation theory, not for this exact label.

Regarding A ← H → C: H is a common cause and, if unobserved, an unobserved confounder. “Unqueried” also allows H to be observed but excluded from the audit. Appendix D.1 explicitly explains the marginalisation of variables outside the queried set.

The assumptions and family definition were already present in the original manuscript; the revision expands their explanation. We respectfully ask that these passages be considered when reassessing the concern about our understanding of established causal concepts. We remain happy to clarify any specific definition or inference.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.