acceptodds
Under review as a conference paper at ICLR 2027

The Relativity of Causal Knowledge: Backbone Identifiability under Private Knowledge

Abstract

Complex causal models may contain far more detail than a shared task requires. A Markov kernel can retain the outcome distributions for selected interventions while omitting other variables and mechanisms. The Relativity of Causal Knowledge (RCK) framework organises communication of such distilled causal quantities along network paths, using shared, interventionally consistent abstractions at each step. This offers the potential to reduce communication demands and disclosure of models and data. We examine an identification requirement for realising that potential. For two causes of a common outcome, we establish local non-identifiability for general kernels under explicit conditions: complete separate intervention reports can leave the joint response undetermined. Known identification obstacles therefore persist in the abstract quantities, even when RCK transport is well defined. Under additive separability, identified response functions determine mean changes; the noise distribution additionally determines the full kernel. We thereby take a step towards operational RCK, highlighting the need to connect its communication framework with causal identification and distributed inference.

Primary area:
causal reasoning
Keywords:
Relative Causal Knowledge, Identifiability

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_rS63OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies whether the two agents’ complete single-cause intervention-response kernels, together with specified background averaging laws, identify a shared RCK response kernel. Under A1–A3, it constructs distinct kernels arbitrarily close to each compatible baseline satisfying those conditions, preserving both complete reports while changing some joint-intervention distributions. Under additive separability and finite first moments, the report means and known background laws identify the joint-intervention mean. Within this additive model, a specified or identified disturbance distribution also determines the full response kernel.

Critical strengths
  1. The local non-identification result goes beyond an isolated counterexample. The construction applies to every baseline satisfying A1–A3 and keeps sufficiently small perturbations in both directions valid. Preservation of the complete reports and arbitrary closeness to the baseline are both established explicitly.

  2. The paper clearly distinguishes identification of means, full distributions, and particular policy queries. It does not equate a nonunique kernel with ambiguity in every query, and it treats separability as an additional sufficient assumption.

Critical weaknesses

The local construction applies to every compatible baseline satisfying A1–A3 and goes beyond an isolated counterexample. My reservation concerns the significance of this extension. Non-identification of joint effects from separate interventions is established, and the paper adds a local characterization under explicit sufficient conditions. This is a useful clarification, but I am not fully convinced that the theoretical advance is substantial enough to support acceptance. My concern is about significance rather than correctness.

Questions for authors

Given the two complete report kernels and the known background laws, are there simple report-level conditions that certify the existence of a compatible baseline satisfying A1–A3? An illustration in the finite-state setting would help clarify how the sufficient conditions can be assessed without knowing the baseline.

3 comments

Thank you for recognising the local construction of the theorem and for asking how its assumptions can be assessed from reports. We added a report-level certificate in Appendix C.1 and clarified the contribution in Section 4 and Appendix E (lines 1534-1561). All manuscript changes are highlighted in blue.

Critical weakness: significance beyond established joint-effect non-identification.

We agree that non-identification from separate interventions is established. Theorem 1 gives a more specific characterisation: on general measurable spaces, it constructs a report-preserving perturbation through every compatible baseline under A1-A3. Both signs of a sufficiently small perturbation remain valid. Restricting attention to nearby candidates therefore leaves the ambiguity, and the proof identifies joint-intervention probabilities that change while both complete reports remain fixed.

For the communication problem motivated by RCK, this distinguishes two requirements: translating available causal information consistently, and having enough information (from the sender or otherwise) to identify the recipient’s query. Our theorem addresses the latter within the stated kernel compatibility class; it does not establish that every perturbation preserves an arbitrary fixed collection of local SCMs and abstraction maps.

The broader motivation is that RCK could support cumulative causal reasoning across researchers or institutions while detailed mechanisms and data remain local. Our recovery result makes this concrete: under the additive model, report means determine relative shifts, so one identified joint-intervention outcome law fixes the full kernel (Section 5; Appendix D, Eq. (41)). With valid RCK maps and a sender value restricting to that law, one additional causal message can answer every joint-intervention query. Acquiring the law still requires additional causal evidence. We included these practical implications in Appendix E (lines 1654-1664).

Question: can the reports certify existence of a baseline satisfying A1-A3?

Yes: finite settings admit simple sufficient conditions. We expanded the existing binary witness in Appendix C.1. With equal background weights, write ri,sjr_i,s_j for the two agents' reported success probabilities. First check their overall averages agree, then construct

\qquad p^{(0)}_{ij}=r_i+s_j-\bar p.$$ If every entry lies strictly between zero and one, this table and the cause backgrounds verify A1-A3. Its row and column averages reproduce the reports (A1); contrasts $(1,-1)$ on each cause are centred (A2); strictly positive success and failure probabilities across all cells provide a common nontrivial outcome component (A3). For example, $r=s=(0.3,0.5)$ gives $\bar p=0.4$ and $$p^{(0)}= \begin{pmatrix}0.2&0.4\\\\0.4&0.6\end{pmatrix}, \qquad p(t)=p^{(0)}+ t\begin{pmatrix}1&-1\\\\-1&1\end{pmatrix}, \quad |t|<0.2.$$ These are distinct valid joint-response tables with exactly the same complete binary-outcome reports. The certificate uses only the reports and the known background weights in Section 4. It certifies existence of a qualifying baseline; the constructed table need not be the true response. Failure of this construction is inconclusive, since another compatible baseline may satisfy A3. For larger finite spaces and unequal known weights, impose the report equations and probability normalisation in a linear programme, then maximise a common lower bound on the cell probabilities. A positive bound certifies A1-A3 when each cause has at least two positive-weight values and there are at least two outcome states. The bound can also be restricted to a cause rectangle with two positive-weight values per cause and at least two outcome states, reflecting local overlap. We added the binary certificate, numerical example and finite feasibility formulation to Appendix C.1, Eqs. (39)-(40). We remain available for further clarifications.
Reviewer_rS63OpenReview ↗

Follow-up after rebuttal

The added finite-state example and linear-programming certificate address my related question. My reservation about the extent of the contribution remains, so I retain my rating of 2.

Thank you for confirming that the finite-state example and linear-programming certificate address your question.

We see the significance in advancing RCK’s goal of causal communication through abstractions while full SCMs remain private. Our latest reply to RcEQ explains how the kernel results and the strengthened RCK constructions together establish the identification limits of this exchange and support progress towards operationalising RCK.

We appreciate your careful engagement and remain available for further questions or feedback.

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

The Authors investigate the identifiability aspects related to Relative Causal Knowledge (RCK), recently formalized in causal modeling through network sheaves and cosheaves of causal knowledge. Let ρ\rho and σ\sigma be two nodes connected by an edge τ\tau. Let CK(Mρ)\mathrm{CK}(M^\rho), CK(Mσ)\mathrm{CK}(M^\sigma), CK(Mτ)\mathrm{CK}(M^\tau) be the causal knowledge induced by three structural causal models (SCMs) MρM^\rho, MσM^\sigma, and MτM^\tau attached to those nodes and edge of three causal agents. MτM^\tau is a common abstract backbone of both MρM^\rho and MσM^\sigma. In plain words, in the considered pairwise setting, the causal knowledge of ρ\rho from the perspective of σ\sigma transported (and thus filtered out) through the common backbone τ\tau (and vice versa) is the RCK. Focusing on the interventional regime, the Authors use Markov kernels to map interventions on the endogenous to interventional measures over a common outcome. The two nodes share the common outcome XX while holding private causes ZρZ_\rho and ZσZ_\sigma; the backbone has both. The local report of an agent, say ρ\rho, is the single-cause interventional kernel χ~ρ(⋅∣zρ)\tilde{\chi}^\rho(\cdot \mid z_\rho), related to the backbone kernel QQ (jointly interventional) via the averaging over the other cause (i.e., ZσZ_\sigma) with fixed and known background law. The paper investigates when the latter backbone kernel on the edge τ\tau is uniquely identified from the single-cause interventional kernels of the agents ρ\rho and σ\sigma living on the nodes. The Authors establish non-identifiability of the common backbone kernel under assumptions (A1)-(A3). Additionally, under additive separability, the identification and communication of the response functions and noise disturbances are sufficient conditions to recover the kernel.

Critical strengths

The usage of Markov kernels to investigate RCK is an interesting approach to manipulate the interventional layer of the causal knowledge. Thm. 1 leverages this machinery to establish the result, whose correctness is not questioned by my assessment. This is consequential for the paper soundness and distinguishes the work from a “clear reject”.

Critical weaknesses
  • The non-identifiability result is established in the kernel compatibility class, but is claimed on the RCK backbone. The result concerns the joint backbone kernel, which is only constrained to be compatible according to Eq. (2). Proposition 2 then invites to read it on the backbone values χτ\chi^\tau, but the constraints remain those in Eq. (2), while nothing, (A1)-(A3) included, concerns the section condition in lines 766-768, which is actually the constraint that must hold on the interventional measures. Nothing establishes (i) that the perturbed kernels QtQ_t reflect the same section condition, (ii) that they are representative of the causal knowledge of the same MτM^\tau, or (iii) that the interventionally consistent causal abstraction relations with MρM^\rho and MσM^\sigma still hold for them. The Authors do qualify the scope in the Appendix (lines 795-796; “Compatibility in RCK notation”, lines 798-809), but those qualifications do not reach the statement of Proposition 2, nor the claim that the result holds even when interventionally consistent abstraction maps are available (lines 21-22, lines 102-103). This is consequential because the RCK reading is the contribution the paper advertises, and it is precisely the part that seems to me not established.
  • The sheaf plays no role in the results. None of the assumptions, theoretical statements, proofs, or examples in the main body explicitly refers to the restriction and extension maps, which from the categorical standpoint are first-class objects of the sheaf; and the topology considered is trivial (two nodes, one edge, and the corresponding incidence relations), whereas in the cited references on sheaves of CK and causal abstraction networks the network topology and the related operators (sheaf and algebraic Laplacian) are key for the convergence to globally coherent states. This is consequential because it determines what remains once the claim is restricted to its correct scope. It is also not clear whether the results would survive on non-trivial networks, where compatibility constraints become inherently coupled.

Update after the discussion

Point 1 has been acknowledged and addressed by the Authors; claims are now consistently qualified with the results. Point 2, on the role of sheaf topology, remains largely open, although some effort has been made in this direction. After manuscript verification, I raised my initial score from 2 to 3.

Questions for authors

Please refer to the “Critical Weaknesses” section. In addition, the following questions could affect my assessment.

  1. In which sense is MτM^\tau an abstraction of MρM^\rho and MσM^\sigma? As conveyed in Example 2 and in the definition of the candidate backbone joint kernel QQ (line 218), MτM^\tau sees both local causes ZρZ_\rho and ZσZ_\sigma, and can therefore perform joint interventions: the edge object is finer than either endpoint, whereas an abstraction is coarser than the model it abstracts. The setting seems closer to the “embedding” direction of causal abstraction networks investigated in the cited reference (D’Acunto et al., 2026), where the model at the edge stalk has the same granularity as the finer model at its endpoints. Relatedly, which of the restriction and extension maps carries the interventionally consistent causal abstraction in your construction?
  2. Where and how does communication occur? In line 427 the Authors say “Agent ρ\rho sends its response kernel χ~ρ\tilde{\chi}^\rho to Agent σ\sigma” (similarly in Example 5, lines 436-437). Is this communication intended to occur on the sheaf, that is, does it correspond to the communication of interventional measures shaped by the action of the restriction and extension maps? The same question applies to the disturbance law. Relatedly, how are the projection operators in line 233 related to the sheaf?
  3. In the RCK framework, agents may not agree at all on the network, that is, the space of global sections may be empty. This is not a failure of the framework, but rather a global property of the system described by it. In a realistic network scenario, do the results implicitly assume that the agents expose to the network their causal belief without breaking any local rule, that is, that the 0-cochain is a global section and the disagreement on the edge stalk vanishes?

The following are comments that do not affect my assessment.

  • Interventional consistency in lines 65-66 is ambiguous. It should be clear that interventional consistency is a property of the causal abstraction, also under a certain correspondence ω\omega that maps an intervention ii at the low-level model into ω(i)\omega(i) at the high-level one.
  • Lines 228-229 are ambiguous. The report is a single-cause interventional kernel, so it is its output χ~ρ(⋅∣zρ)\tilde{\chi}^\rho(\cdot \mid z_\rho) that is assumed to be a valid interventional measure.
  • In lines 238-239, πρ\pi^\rho and πσ\pi^\sigma are said to be local report operators. In reality, what is local is their output; their definition in line 233 is based on joint information that is not locally available to the agents.
  • “Admissible” in (A1) should be “compatible”.
5 comments

Thank you for identifying the gap in Proposition 2 and your questions on abstraction, communication, sections and topology. You're right: A1-A3 give report-preserving alternatives but our RCK claim also requires causal membership, sections and IC abstractions. We uploaded an updated paper: revised Prop 2 and proof sketch are in Sec 4; full proof in App B.2. All changes are highlighted in blue.

Q1/W1: backbone and RCK claim.

The revised construction uses a coarser backbone: each local SCM MaM^a contains its cause, latent context and outcome; MτM^\tau contains only XX. Its causal space CK(Mτ)CK(M^\tau) is the edge stalk. The recipient's query is

$ \chi_t^\tau(z,n)=Q_t(\cdot\mid z,n) =P_{M_t}(X\in\cdot\mid\mathrm{do}(Z=z,N=n)). $

Here (z,n)(z,n) index joint interventions in MtM_t; the backbone represents their outcome laws. The endogenous map φa(za,c,x)=x\varphi_a(z_a,c,x)=x induces the restriction \alpha_a=(\varphi_a)_\\# along the IC abstraction Ma→MτM^a\to M^\tau. The extension βa\beta_a returns a local causal value with αa(βa(χ))=χ\alpha_a(\beta_a(\chi))=\chi.

Revised Proposition 2 gives a constructive existence proof: for any nondegenerate real-valued cause backgrounds, it constructs fixed RCK components and a continuum of joint SCMs giving different answers to the same joint-intervention query despite identical recipient information (Eq. (5)): Dσ:=(Mσ,(αa,βa)a∈{ρ,σ},χ~ρ,χ~σ).\mathcal D_\sigma:=(M^\sigma,(\alpha_a,\beta_a)_{a\in\{\rho,\sigma\}},\tilde\chi^\rho,\tilde\chi^\sigma). Specifically, we prove four new lemmas:

  • Lemma 9 adapts existing Propositions 6-7 to construct Eq. (4)'s family within joint SCMs, preserving complete local SCMs and reports. It verifies A1 and A3 for Q0Q_0, and A2 for u,vu,v.

  • Lemma 10 proves fixed IC abstractions Mt→Ma→MτM_t\to M^a\to M^\tau: state/noise maps, noise pushforwards and corresponding hard and parent-preserving soft interventions. It supplies affine restrictions and extensions with αa∘βa=id\alpha_a\circ\beta_a=\mathrm{id}, and shows that each βa(χ)∈CK(Ma)\beta_a(\chi)\in CK(M^a) is induced by a soft intervention of the same fixed local SCM MaM^a (Eq. (33)).

  • Lemma 11 proves that the projected joint-query laws χta(z,n)\chi_t^a(z,n) belong to CK(Ma)CK(M^a) and satisfy the section condition (2): αρ(χtρ(z,n))=Qt(⋅∣z,n)=ασ(χtσ(z,n)).\alpha_\rho(\chi_t^\rho(z,n))=Q_t(\cdot\mid z,n)=\alpha_\sigma(\chi_t^\sigma(z,n)). Each candidate answer is realised by a soft intervention of the same backbone. This intervention varies with tt when u(z)v(n)≠0u(z)v(n)\ne0; the original joint intervention remains fixed.

  • Lemma 12, using Proposition 8, proves that the joint outcome laws and means differ with tt wherever u(z)v(n)≠0u(z)v(n)\ne0, while Dσ\mathcal D_\sigma is unchanged.

Adapting Theorem 1's perturbation to fixed RCK components required additional construction assumptions: mutually independent cause roots, uniform latent contexts and atomless scalar outcome noise, with a positive density adapted to two commuting, reference-law-preserving, fixed-point-free involutions. Each reverses the perturbation SS. These choices give invertible reference maps and preserve local noise laws under perturbation. Theorem 1 retains its general report-level scope.

Q2: communication and report generation.

The reports encode each agent's initially private knowledge of intervention responses. RCK transports a sender value as βσ(αρ(χρ))\beta_\sigma(\alpha_\rho(\chi^\rho)) (D'Acunto and Battiloro, 2025, Definition 11). In our construction, restriction gives the reported outcome law (Lemma 10, Eq. (31)):

$ \pi^a(Q_t)(\cdot\mid z_a) =\alpha_a(P_{M^a}(\cdot\mid\mathrm{do}(Z_a=z_a))) =\tilde\chi^a(\cdot\mid z_a). $

πa\pi^a acts on joint kernels; αa\alpha_a acts on local causal values. This equality holds for our construction; RCK does not require this report representation (Appendix E, lines 1562-1576).

Under the additive model, the reports give relative shifts; one identified joint-response law then determines the full kernel (Section 5; Appendix D, Eq. (41)). A known absolute shift also yields PεP_\varepsilon. RCK maps transport a sender value restricting to that law. Acquiring it requires additional causal evidence beyond our local reports.

Q3/W2: sections and topology.

Yes: a global section entails agreement after restriction (RCK, Definition 10). Lemma 11 constructs a query section for each candidate tt. Agreement holds within each section, while different sections give different joint-query answers compatible with the same Dσ\mathcal D_\sigma. The query values χta(z,n)\chi_t^a(z,n) are absent from Dσ\mathcal D_\sigma; supplying a particular query section would fix its edge value.

We clarified sections and the two-node scope in Appendix E (lines 1577-1588). Additional edges can couple constraints and remove ambiguity; the example extends to networks whose added edges give no distinguishing constraint. Topology-dependent identification and convergence require further study.

Minor comments addressed in the revised paper.

We welcome your feedback on this construction and qualified scope.

Reviewer_RcEQOpenReview ↗

Response to the Authors

I sincerely thank the Authors for addressing my concerns and for promptly providing a revised version of the manuscript, which aids in the reviewing process. Please find below my comments on the Authors’ response.

[Q1/W1] Thaks for acknowledging the gap in the original Prop. 2. To better understand, I would appreciate some clarification on the following points:

  • If I understand correctly, QtQ_t is the kernel of MtM_t rather than of the backbone (“Here QQ collects the laws indexed by interventions in the joint SCMs”, line 237). If so, how should the references to a backbone kernel in Thm 1 be read, together with the claims in lines 21-22 and 102-103?
  • From my understanding of Lem. 11, the joint intervention do(Z=z,N=n)\mathrm{do}(Z=z, N=n) is not paired by the correspondence with a fixed intervention of MρM^\rho, MσM^\sigma or MτM^\tau. The construction instead starts from the law Qt(⋅∣z,n)Q_t(\cdot \mid z,n) induced by MtM_t, and then looks for interventions at the local and backbone levels that reproduce it. My concern is that, in this way, membership in CK(Mρ)CK(M^\rho), CK(Mσ)CK(M^\sigma) and CK(Mτ)CK(M^\tau) depends on the expressivity of these SCMs, and not on the IC abstraction relations. In more detail: (i) the backbone realizes any law on X\mathcal{X}, and the same holds for the local outcome laws; (ii) both restrictions select the common outcome of the same joint law (Eq. (36)). Hence, membership and the section condition seem to hold for every candidate, and do not restrict the candidate answers. I would ask the Authors to clarify what the RCK layer of Prop. 2 adds to Thm. 1, beyond showing that suitable RCK maps do not exclude the ambiguity. If nothing is added, the claims in lines 21-22 and 102-103 should be qualified accordingly.

[Q2] In Prop. 2, and based on my understanding of Lem. 11, a local value restricting to the joint-intervention law Qt(⋅∣z⋆,n⋆)Q_t(\cdot \mid z^\star, n^\star) is realized only through an intervention chosen according to tt. This suggests that local SCMs need not contain such a value. With reference to the paragraph “Recovery from one joint-intervention law” in App.D (starting at line 1492), which agent would hold such an identified joint-intervention law, and how would it be obtained at the local SCM level?

[Q3/W2] I believe the two-node scope of the investigation is important for the reader, especially because RCK is presented as a framework organizing communication across networks (e.g., lines 15-17, 36-39). I therefore encourage the Authors not to confine this aspect to App. E, and to state it explicitly in the main text. As far as the extension to larger networks is concerned, the example provided by the Authors in lines 1585-1587 uses copies of a local SCM linked by identity maps, which add no information by construction. A non-degenerate example, in which additional edges carry information and the ambiguity nonetheless persists, would strengthen the paper.

We truly appreciate the time and care devoted to this exchange. Your feedback has helped us sharpen the claims and strengthen the paper.

Q1/W1: joint-response kernels and the information gap.

1. Kernel and backbone. QtQ_t is MtM_t's joint-response kernel (Eq. (23)). Prop. 2 represents each answer as Qt(⋅∣z,n)=χtτ(z,n)∈CK(Mτ)Q_t(\cdot\mid z,n)=\chi_t^\tau(z,n)\in CK(M^\tau) on XX's state space (Eq. (36)); (z,n)(z,n) specifies interventions on both private causes. We will call Theorem 1's object a “joint-response kernel” and explain this link in Section 4.

2. Expressivity and missing information. Sender ρ\rho supplies responses over ZZ; recipient σ\sigma has responses over NN and its complete SCM. Each report averages the other cause. RCK transports the sender's law (Eqs. (30)–(31)), while the recipient asks for the response at both Z=z,N=nZ=z,N=n. The interaction term tu(z)v(n)Stu(z)v(n)S cancels in both reports yet changes that answer (Eq. (4)).

You are right: expressivity realises each candidate in CK(Mρ)CK(M^\rho), CK(Mσ)CK(M^\sigma) and CK(Mτ)CK(M^\tau); selecting the common outcome gives the section condition (Lemma 11). Prop. 2 proves that the information gap persists with the same Dσ=(Mσ,(αa,βa)a∈{ρ,σ},χ~ρ,χ~σ)\mathcal D_\sigma=(M^\sigma,(\alpha_a,\beta_a)_{a\in\{\rho,\sigma\}},\tilde\chi^\rho,\tilde\chi^\sigma): complete local SCMs and valid IC maps stay fixed (Lemmas 9–10), yet sections give different answers to the same query (Lemma 12). It is an existence result under the construction assumptions.

We will explain this after Prop. 2 in Section 4 and in Appendix E, and replace:

  • Lines 21–22: “We realise this ambiguity for joint-intervention outcome laws represented on shared backbones, while complete local SCMs and RCK communication maps remain fixed.”
  • Lines 102–103: “Proposition 2 constructs a two-agent system with fixed complete local SCMs and valid RCK maps: the same private reports admit different joint-SCM response laws. Each candidate answer is represented as a causal value of the same outcome-only backbone.”

Q2: who holds the law, and how is it realised locally?

In Prop. 2, each fixed local SCM contains a causal value mapping through the same fixed restriction αa\alpha_a to Qt(⋅∣z⋆,n⋆)Q_t(\cdot\mid z^\star,n^\star), for every tt and pair (z⋆,n⋆)(z^\star,n^\star) (Lemma 11, Eq. (36)). The realising intervention varies with its target law. Eqs. (32)–(33) realise an identified law's extension using that law and fixed noise.

In Appendix D, either agent independently identifies χ⋆τ=Q(⋅∣z⋆,n⋆)\chi_\star^\tau=Q(\cdot\mid z^\star,n^\star) from a study or an experiment where each sets its cause. If the sender holds it, its local value βρ(χ⋆τ)\beta_\rho(\chi_\star^\tau) is sent as βσ(χ⋆τ)\beta_\sigma(\chi_\star^\tau); restriction recovers the law. If the recipient holds it, recovery is direct. General SCMs require admissibility and causal extensions.

Under structural additivity, an intervention-invariant disturbance law and finite means, Eq. (41) centres the supplied law and adds the joint mean from the reports and the recipient's own cause law, recovering the full kernel without starred levels or tt. We will clarify Section 5 and Appendix D.

Q3/W2: network scope and informative edges.

We will state Prop. 2's two-agent scope and add a non-degenerate three-agent extension in Section 4, with details in Appendix E.1.

Add sender γ\gamma to form ρ∼σ∼γ\rho\sim\sigma\sim\gamma. All agents use Lemma 9's (cause, context, outcome) SCMs: ρ:(Z,Cρ,Xρ)\rho:(Z,C_\rho,X_\rho), σ:(N,Cσ,Xσ)\sigma:(N,C_\sigma,X_\sigma), γ:(Zγ,Cγ,Xγ)\gamma:(Z_\gamma,C_\gamma,X_\gamma). The backbones carry Xτ=XX^\tau=X on ρ∼σ\rho\sim\sigma and coarser Xτ′=ψ(Xτ)X^{\tau'}=\psi(X^\tau) on σ∼γ\sigma\sim\gamma; both σ,γ\sigma,\gamma use the fixed many-to-one map ψ\psi.

Add a small term with coefficient textt_{\mathrm{ext}} to Prop. 2's scalar outcome-noise density. It cancels under Eq. (27)'s averaging, preserving complete local SCMs and reports. The joint-query kernel Qt,textQ_{t,t_{\mathrm{ext}}} coarsens to KtK_t on σ∼γ\sigma\sim\gamma: ψ\psi removes only the new perturbation. We will extend Lemmas 10–11's IC and causal-extension proofs, including parent preservation, with all maps fixed across both parameters.

Sender γ\gamma supplies an independently identified law ζ⋆=Ktˉ(⋅∣z⋆,n⋆)\zeta_\star=K_{\bar t}(\cdot\mid z^\star,n^\star) for the query on Z,NZ,N, with both levels known to γ,σ\gamma,\sigma and u(z⋆)v(n⋆)≠0u(z^\star)v(n^\star)\ne0. Within this family, the law and levels identify tˉ\bar t (Eq. (38)). Recipient σ\sigma also has its own report and ρ\rho's. Agreement on σ∼γ\sigma\sim\gamma fixes t=tˉt=\bar t. Each admissible textt_{\mathrm{ext}} admits a global section agreeing with γ\gamma, yet an event for XX has probability 14[1+(tˉ+text)u(z⋆)v(n⋆)].\tfrac14[1+(\bar t+t_{\mathrm{ext}})u(z^\star)v(n^\star)]. Thus the full kernel remains unidentified from this information, even though the additional sender's message narrows the compatible family.

General criteria across topologies, maps and supplied information remain open.

We are happy to upload a revised manuscript if helpful. Thanks again.

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

This is a theory paper about how what kind/how much information needs to be shared in order to jointly reason about some common causal system. The paper provides a general unidentifability result and then an explicit assumption that enables identifiability. The paper includes a running example throughout to make the theoretical results more concrete.

Critical strengths
  • The paper rigorously addresses a fundamental and pervasive problem: how do we use data we have access to in order to reason about underlying causal systems that we want to understand (in particular, when we have local/private data and want to understand a shared causal abstraction)
Critical weaknesses
  • As a theoretical paper in a fairly niche topic, without an implementation or demonstrated real data application (or even the goal of learning something), it's impact could be limited. (at the same time, I recognize that this theory provides a rigorous foundation for future work to do learning)
Questions for authors
  1. How practical/testable is additive separability?
  2. What stands between the results in this paper and a real application on, e.g., the kind of data in the running example?
3 comments

Thank you for the constructive questions about separability and application. They helped us specify the evidence needed for recovery. All manuscript changes are highlighted in blue. We added an assessment criterion in Section 5 and application requirements in Appendix E (lines 1654-1664).

Q1: how practical or testable is additive separability?

At the level of expected outcomes, separability means that the effect of changing one cause does not depend on the other's level. In our education example, a teacher-quality improvement would have the same effect on expected earnings across neighbourhood exposures. This is a substantive causal assumption.

It can be assessed using identified joint-intervention responses. Write mˉ(z,n)\bar m(z,n) for expected earnings when teacher and neighbourhood exposures are set to z,nz,n. At two levels of each exposure, additive means imply (Section 5, Eq. (8))

mˉ(z1,n1)−mˉ(z1,n0)−mˉ(z0,n1)+mˉ(z0,n0)=0.\bar m(z_1,n_1)-\bar m(z_1,n_0)-\bar m(z_0,n_1)+\bar m(z_0,n_0)=0.

A nonzero population contrast rules out mean additivity. A practical test would estimate the four joint means and assess whether the contrast differs from zero beyond sampling uncertainty. This requires a coordinated experiment, in which each agent sets its own cause, or another justified identification method. Passing this check at selected levels leaves separability at other levels unresolved.

Under the paper's additive structural equation (Eq. (6)), changing either cause while holding the other fixed shifts the outcome distribution without changing its shape. Under the fixed averaging laws, each agent's reports must also differ only by shifts. A change in variance, for example, contradicts Eq. (6). Yet reports differing only by shifts can hide joint interactions: these can cancel when averaging over the other cause (Appendix C.3).

The revised Proposition 2 illustrates what additional information can suffice: within its specified family, with only the interaction strength unknown, one identified joint-response probability at suitable known intervention levels determines that strength and hence the joint responses (Eq. (38); Appendix E, lines 1708-1710). This sufficiency relies on restricting the remaining uncertainty to one parameter. The required evidence could come from an existing study, justified observational identification or a targeted experiment. RCK supports exchanging available causal responses while keeping detailed local SCMs private.

Separability is useful where domain knowledge supports it and joint-response evidence can assess its implications. Section 5 now states testable restrictions and their data requirements. Developing and evaluating statistical tests, including their calibration and power, remains future work.

Q2 and critical weakness: what stands between these results and a real application?

We agree that an implementation and a validated application are important next steps. Our results establish which information suffices before introducing estimation error.

A real application requires:

  1. A common interface with aligned outcomes, intervention meanings and target populations, supported by justified IC abstraction maps.
  2. Causally identified reports, using experimental or observational identification methods, with uncertainty estimates.
  3. Valid averaging relations: the background laws must describe the intended regimes. If an intervention changes the other cause's distribution, that change must be modelled.
  4. Identification of the policy query from the recipient's local knowledge, received reports and any additional evidence. Recovery assumptions, such as separability, need support from domain knowledge and assessment against causal evidence where available.

The paper gives sufficient conditions for different queries. Mean additivity, response contrasts and a reference mean suffice for compensation. For full-kernel recovery, one sufficient route is the structural additive model together with the reports and one identified joint-intervention outcome law (Appendix D, Eq. (41)).

Example 5 illustrates the calculation under these conditions; Appendix E (lines 1596-1669) details the worked interface. A teacher effect of 1.3 percentage points per standard-deviation year would offset a 9-percentage-point earnings shortfall with approximately 6.9 standard-deviation years, or 0.53 per year over thirteen years. This remains an illustrative conditional calculation. The cited studies do not themselves establish the aligned report kernels and abstraction maps required for this exchange.

Combining causal knowledge could let researchers answer shared questions while keeping detailed SCMs and data local. Targeted joint evidence may resolve a policy query without identifying the entire kernel. We updated Section 5 and Appendix E (lines 1654-1664) to include these points, with the changes highlighted in blue.

We remain available for further clarifications.

Thanks for the thorough rebuttal. I think the edits improve the paper, and I maintain my positive rating.

Thanks for the encouraging response. We are pleased to hear that our revisions have improved the paper, and we sincerely appreciate the time and care you have devoted to reviewing our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.