acceptodds
Under review as a conference paper at ICLR 2027

Characterizing Constructive Causal Abstraction via Intervention Composition

Abstract

Causal abstraction relates causal models across levels of description. Constructive abstractions are widely used and provide useful structural properties. Beckers and Halpern conjectured that strong -abstraction would imply constructivity under minor technical conditions. Existing positive results concern linear structural causal models under additional assumptions. We address the general case. First, a fixed edgeless low-level model admits strong abstractions with arbitrary high-level DAGs, even with analytic mechanisms, independent full-support noise, and a full-rank abstraction mapping. Second, for strong -abstractions, preservation of intervention composition on the full canonical domain is sufficient for constructivity. This condition is stronger than constructivity of the mapping: natural aggregations admit canonical interventions that violate composition preservation. Such interventions also cause failures in published abstraction results. We restrict the low-level domain while preserving every realizable high-level intervention. For a surjective mapping with full high-level intervention coverage, composition preservation on either restricted domain is then equivalent to constructivity. Finally, a controlled nonlinear learning experiment uses composition constraints to improve prediction over unregularized learning when input correlations change.

Primary area:
causal reasoning
Keywords:
causal abstraction, constructive abstraction, structural causal models, intervention composition, causal representation learning, causal inference

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_CK2pOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper takes two causal abstraction definitions by Beckers and Halpern in the paper Abstracting Causal Models and examines their relationship in general, non-linear settings.

Strong τ\tau-abstraction, the weaker of the two definitions in the Beckers and Halpern paper, requires that for every possible high-level intervention, there exist compatible, potentially overlapping low-level interventions. Constructive τ\tau-abstraction, in addition, requires an injection from low-level micro-variables to high-level macro-variables. Beckers and Halpern predict that constructive abstractions as the ones which will be the most practical, and left the "few minor technical conditions" to go from strong -> constructive to future work.

This paper claims to have the following theoretical contributions

  1. A counter-example that shows that strong does not implies constructive, under some regularity conditions.
  2. A proof that strong + intervention-composition-preserving on ILτ\mathcal I^\tau_L => constructive, where ILτ\mathcal I^\tau_L is the set of all low-level interventions with high-level counterparts.
  3. A proof that strong + intervention-composition-preserving on Jτrel\mathcal J^{rel}_\tau <=> constructive, where Jτrel⊂ILτ\mathcal J^{rel}_\tau \subset \mathcal I^\tau_L is the set of low-level interventions that only contains "relevant" low-level variables. A low-level variable is considered relevant if intervening on it by itself corresponds to an intervention in the high-level model.

The paper then follows up with a very small scale experiment on synthetic data, in which the low-level model with state space R8\mathbb R^8 maps to a high-level model with state space R4\mathbb R^4. Variations of of a penalty term that enforces intervention composition preservation at different levels of the model architecture are introduced as a part of the training loss, and shown to reduce prediction MSE.

Critical strengths

The paper demonstrates that a strong abstraction is a non-trivially weaker definition than constructive abstractions in general settings, and exactly characterizes the conditions to go from strong -> constructive as the property intervention-composition-preservation on a relevant set of low-level interventions -- which disproves Beckers and Halpern's conjecture.

Critical weaknesses
  1. I have some potential doubts about the overall significance of the theoretical results. Beckers and Halpern introduced a sequence of definitions for causal abstractions because they felt that the notion of exact transformations introduced by Rubenstein et al in their paper Causal Consistency of Structural Equation Models was not restrictive enough. The sequence of successively more restrictive definitions ends up being:

exact transformations -> uniform transformations -> abstraction -> strong abstraction -> constructive abstraction

Beckers and Halpern favor the constructive definition as the most useful for practical applications, while conjecturing that the strong definition is very close to it, which motivates the strong definition in the first place. This paper shows that "strong abstraction + intervention compositionality condition = constructive abstraction", which is an interesting result. However, now I'm not sure I see the point in the strong definition, unless it occurs naturally and frequently in practical applications.

  1. The presentation of the paper is poor and almost impossible to read without referring to the background literature. In many places, statements are not rigorous where they ought to be and new notation pops up without being defined first (e.g. the opening sentence of Def 2.5). While I think I can understand the main theoretical claims and their underlying intuitions, I have trouble verifying them properly in the time I have allotted for reviewing this paper.

  2. The experiment doesn't support the main theoretical results. I might have the wrong interpretation, but the fact that the compositionality penalty lowers the predicted MSE in this particular experimental setup seems to have nothing to do with the definitions of strong nor constructive abstractions. In fact the paper itself states "These finite tests ... do not establish exact constructivity of the trained networks".

Questions for authors
  1. Is there a difference between the constructive abstractions given in Def 2.5 of this paper, in comparison to the constructive τ\tau - abstraction of Beckers and Halpern?

  2. How does Def 2.6 relate to the notion of intervention composition given by Lorenz and Tull?

  3. Are there natural settings / applications in which the definition of a strong abstraction is more appropriate than the definition of a constructive abstraction?

  4. The paper Causal Abstractions, Categorically Unified by Englberger and Dhami claims to "describe examples of strong τ -abstractions that are not constructive τ -abstractions" -- how would you defend the novelty of theorem 1 in this paper against their results?

4 comments

Thank you for your thoughtful comments, suggestions and the time you devoted to the review. Below is our response to every weakness and question.


Clarification before response

Thank you for your helpful summary of our paper. We would like to clarify one point about Theorem 3. This may simply be a typo, but we mention it to avoid any possible misunderstanding. Our Theorem 3 is not

strong + intervention-composition-preserving on Jτrel\mathcal{J}_{\tau }^{rel} <=> constructive

Instead, under the assumptions stated in the theorem, it is

intervention-composition-preserving on Jτrel\mathcal{J}_{\tau}^{rel} <=> constructivity of the mapping τ\tau.

The key difference is that Theorem 3 does not require strong abstraction.


Weakness 1

We agree that the practical limitations of strong abstraction deserve attention, and this is part of the motivation for Section 4 in our manuscript. In Section 4, we aim to clarify the limitations of strong abstraction and characterize constructivity on appropriate intervention domains, rather than advocate strong abstraction as a general requirement. Our paper does not recommend that the causal abstraction community continue using strong τ-abstraction and the original definition of constructive abstraction given in Beckers and Halpern's paper. We also do not think that strong abstraction, which requires all compatible low-level interventions to satisfy interventional consistency, “occurs naturally and frequently in practical applications.”

To clarify how this concern relates to our results, we briefly explain the progression of the paper. In Section 3, we first discuss the conjecture using Beckers and Halpern’s definitions of strong abstraction and constructive abstraction. We then find that even the very simple summation case Y = X1 + X2 may violate strong abstraction. Thus, if we use Beckers and Halpern’s definitions, our theoretical results in Section 3 can only apply to limited cases. In Section 4, we therefore move beyond the original strong-abstraction requirement and discuss the conjecture on a more reasonable intervention domain. Section 4.1 illustrates how the original strong-abstraction requirement excludes natural aggregation mappings and motivates considering smaller intervention domains. At the end of Section 4.1, we write, “We therefore recommend stating the intervention domain explicitly when using constructive abstraction” (lines 242–243 in our original manuscript). In Section 4.2, we show that, on a smaller but more reasonable intervention domain, intervention composition preservation is equivalent to constructivity.


Weakness 2

Thank you for pointing out the presentation issue. The opening sentence of Definition 2.5 introduces and defines the notation you mention. We will make this definition more explicit in the revised manuscript.


Weakness 3

The experiment seems unrelated to the definitions of strong or constructive abstraction.

The experiment tests whether the composition condition (equivalent to constructivity) in theorem 3 can provide a useful training constraint. Take Output Comp (Eq. 7) as an example:

LoutComp=13B∑r=1B∑j≠kr[τθ,j(x′(r))−τθ,j(x~(r))]2.\mathcal L_{\mathrm{out}}^{\mathrm{Comp}} = \frac{1}{3B}\sum_{r=1}^{B}\sum_{j\ne k_r} \left[ \tau_{\theta,j}(x'^{(r)}) - \tau_{\theta,j}(\tilde x^{(r)}) \right]^2.

Our theoretical composition condition is

ωτ(i∘k)=ωτ(i)∘ωτ(k).\omega_\tau(i\circ k)=\omega_\tau(i)\circ\omega_\tau(k).

For sample rr, take k=(VL←x~(r))k=(V_L\leftarrow\tilde x^{(r)}), which fixes all low-level variables, and i=((akr,bkr)←(cr,cr))i=((a_{k_r},b_{k_r})\leftarrow(c_r,c_r)), which overwrites one input pair. Their composition fixes the low-level state to x′(r)x'^{(r)}.

The left-hand side therefore fixes all high-level variables to τθ(x′(r))\tau_\theta(x'^{(r)}). The right-hand side first fixes them to τθ(x~(r))\tau_\theta(\tilde x^{(r)}), then overwrites only YkrY_{k_r} with crc_r. Thus, Comp requires

τθ,j(x′(r))=τθ,j(x~(r)),j≠kr.\tau_{\theta,j}(x'^{(r)})=\tau_{\theta,j}(\tilde x^{(r)}), \qquad j\ne k_r.

Therefore we penalize τθ,j(x′(r))−τθ,j(x~(r))\tau_{\theta,j}(x'^{(r)})-\tau_{\theta,j}(\tilde x^{(r)}) in our loss.

In fact the paper itself states "These finite tests ... do not establish exact constructivity of the trained networks".

The sentence you quote does not mean that the experiment is unrelated to constructivity. Under the assumptions of Theorem 3, the necessary and sufficient condition is i∘k∈Di\circ k\in D and ωτ(i∘k)=ωτ(i)∘ωτ(k)\omega_\tau(i\circ k)=\omega_\tau(i)\circ\omega_\tau(k) for all i,k∈Di,k\in D, where D=JτD=\mathcal J_\tau or Jτrel\mathcal J_\tau^{\mathrm{rel}}. In our experiment, intervention values are continuous, so there are infinitely many possible interventions. In practical we cannot test them all, so we use finitely many sampled checks to see whether the resulting training constraint improves prediction.


Question 1

The relationship can be summarized as:

Constructive τ\tau-abstraction in Beckers and Halpern = constructivity of the mapping τ\tau (our Definition 2.5) + strong τ\tau-abstraction.

Our Definition 2.5 isolates the first requirement, constructivity: different high-level variables depend on disjoint sets of low-level variables. Why do we not use Beckers and Halpern's definition of constructive abstraction? Their definition requires strong abstraction, which, as we show in Section 4, excludes many common cases of abstraction. We therefore do not impose this requirement.

And thanks for your question, we will add a brief explanation of this relationship in section 2 of the revised manuscripts for readability.


Question 2

Our definition of intervention composition is the same as Lorenz and Tull's (Example 69(2)): later assignments overwrite earlier ones on shared variables. However, their Proposition 70(a) considers a different setting: it assumes a disjoint variable alignment and studies the induced intervention map on complete aligned blocks. In contrast, our Theorem 3 uses domains defined directly from τ\tau and its canonical intervention map, without assuming constructivity.


Question 3

As discussed in our response to Weakness 1, our contribution does not rely on strong abstraction being more appropriate than constructive abstraction in applications. We study the original strong-abstraction requirement in Section 3 to address Beckers and Halpern’s conjecture. Section 4 then explains why this requirement can exclude natural aggregations and considers smaller intervention domains while preserving full high-level intervention coverage.

In particular, Theorem 3 does not assume strong abstraction: it characterizes constructivity of the mapping through composition preservation on these restricted domains. Corollary 4 applies this characterization to causal models satisfying interventional consistency on the chosen domain. Thus, our practical motivation is to understand when constructivity follows from intervention requirements on an appropriate domain, rather than to identify applications in which the original strong-abstraction definition should be preferred.


Question 4

Englberger and Dhami (2025) examples shows that strong abstraction need not be constructive. Theorem 1 strengthens this observation in two ways.

1. Counterexamples with substantial regularity.

Beckers and Halpern’s conjecture can be summarized as

Strong τ-abstraction+minor technical conditions;⟹;constructivity.\text{Strong }\tau\text{-abstraction} +\text{minor technical conditions} \\;\Longrightarrow\\; \text{constructivity}.

For each mm, Theorem 1 constructs one fixed edgeless low-level SCM that admits every mm-node DAG as a strong τ\tau-abstraction. As explained in the construction in Section 3, whenever the high-level DAG contains an edge, the mapping is non-constructive. Moreover, these constructions satisfy some regularity conditions: analytic mechanisms, independent full-support noise blocks, and analytic, surjective abstraction mappings with full-row-rank Jacobians everywhere. Thus,

Strong τ-abstraction+these regularity conditions;⟹̸;constructivity.\text{Strong }\tau\text{-abstraction} +\text{these regularity conditions} \\;\not\Longrightarrow\\; \text{constructivity}.

This identifies substantial regularity conditions that are insufficient to establish the conjectured implication.

2. A graphical consequence of lacking constructivity.

Schooltink and Zennaro (2025) establish connections between low-level and high-level causal graphs under constructive abstraction. For the fixed edgeless low-level SCM constructed in Theorem 1, the contrast is particularly direct:

Constructive abstraction (Schooltink and Zennaro, 2025):

Edgeless low-level SCM + constructive abstraction ⇒ edgeless high-level DAG.

Strong abstraction (our Theorem 1):

Edgeless low-level SCM + strong abstraction ⇒ arbitrary high-level DAG.

Specifically, under constructive abstraction, our edgeless low-level SCM can only correspond to an edgeless high-level DAG. However, weakening the requirement to strong abstraction allows the same fixed SCM to correspond to any high-level DAG on mm nodes. This illustrates how differently strong and constructive abstraction constrain the correspondence between low- and high-level causal graphs.

Schooltink, W. & Zennaro, F.M.. (2025). Aligning Graphical and Functional Causal Abstractions. CLeaR 2025

Thank you for your detailed responses, in particular the clarification on Thm 3 and your definition of constructivity. My concerns about the practical limitations of strong abstractions (W1) are now resolved, since Thm 3 does not rely on the strong definition.

However, I am now more concerned about the discrepancy between your definition of constructivity and Beckers and Halpern's, as you acknowledged in response to Q1. Given that Beckers and Halpern's definition of constructive causal abstractions is the only widely accepted one in this field, the introduction of a new definition of constructivity should be made very clear in the abstract as well as the introduction, not just with an added sentence in section 2.

This is not to say that I cannot see the motivation behind introducing this new definition, as you have explained in your rebuttal responses. I just think that the framing of this work should be re-assessed. In fact, even the paper title can potentially be seen as misleading, since instead of characterizing the existing definition of constructivity, you re-defined constructivity and shown properties under this new definition.

Additionally, I still maintain that the experiments seem rather superfluous. My concern in W3 was not that "the experiment seems unrelated to the definitions of strong or constructive abstraction", but that the experiments do not validate any of Thms 1-3. Testing whether the composition condition can provide a useful training constraint is an interesting but separate question.

We thank the reviewer for saying that we have addressed Weakness 1. And we want to clarify the point the reviewer is most concerned about because we think there might still be a misunderstanding.

“Given that Beckers and Halpern's definition of constructive causal abstractions is the only widely accepted one in this field, the introduction of a new definition of constructivity […]”

We did not define a new definition of constructivity. As we already said in our previous rebuttal, Beckers and Halpern’s definition of constructive τ\tau-abstraction can be divided into two parts:

Constructive τ\tau-abstraction in Beckers and Halpern = constructivity of the mapping τ\tau (our Definition 2.5) + strong τ\tau-abstraction.

Of these two parts, strong τ\tau-abstraction is just a condition on the relation between the two causal models. The concept of “constructivity” is reflected only in the constructivity of the mapping (i.e., the relevant sets are pairwise disjoint: Rj∩Rk=∅R_j \cap R_k = \varnothing for all j≠kj \ne k). See Beckers, Eberhardt, and Halpern, Approximate Causal Abstractions, Definition 2.6. There they explicitly say what it means for τ\tau to be constructive, and this definition does not depend on strong τ\tau-abstraction.

And we borrow exactly the same definition of constructivity of the mapping.


“the paper title can potentially be seen as misleading, since instead of characterizing the existing definition of constructivity, you re-defined constructivity and shown properties under this new definition.”

As we have said, this is not a “new definition” of constructivity. And we will show more specifically why our title, Characterizing Constructive Causal Abstraction via Intervention Composition, is not misleading even for the readers who only want to see the characterization of constructive tau abstraction in Beckers and Halpern.

Assume there is a reader who likes strong abstraction very much and only cares about Beckers and Halpern’s definition of constructive τ\tau-abstraction, which requires strong τ\tau-abstraction. Theorems 1 and 2 follow Beckers and Halpern’s definition of constructive τ\tau-abstraction exactly, including the strong τ\tau-abstraction condition. For Theorem 2, we assume strong τ\tau-abstraction and show that preserving intervention composition on the full canonical domain implies constructive τ\tau-abstraction in their sense. And for Theorem 3, the fact that it does not rely on strong τ\tau-abstraction does not mean that the result does not hold under strong τ\tau-abstraction. If we simply add the strong τ\tau-abstraction condition to Theorem 3 then we have

Under strong τ\tau-abstraction between M_L and M_H:

ωτ\omega_{\tau} preserves intervention composition on Jτ\mathcal{J}_{\tau} or Jτrel\mathcal{J}_{\tau}^{\mathrm{rel}} ⟺\Longleftrightarrow M_H is a constructive τ-abstraction of M_L in Beckers and Halpern’s definition

The proof is trivial because our original proof of Theorem 3 does not rely on strong τ\tau-abstraction, and none of its assumptions conflict with strong τ\tau-abstraction.

Therefore, this reader gets an exact characterization of constructivity under strong τ\tau-abstraction.

My concern in W3 was not that "the experiment seems unrelated to the definitions of strong or constructive abstraction", but that the experiments do not validate any of Thms 1-3.

We thank the reviewer for emphasizing again the need to make clearer the purpose and limitation of the current experiments, which we tried to do in the updated manuscript, also in response to the other reviewer's comments. Importantly, the purpose of the experiment is not to validate any of the theorems, for as theorems, they need no validation, but rather to demonstrate the potential of designing new learning objectives inspired by the characterization. We agree that the attempt is still very limited and have made this clear in the revised paper.

Reviewer_A4s3OpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper shows that strong abstraction together with intervention composition forces constructive abstractions. The paper gives sufficient conditions for constructive abstractions and shows that when the intervention domain is restricted these conditions become necessary.

Critical strengths

This paper gives theoretical results about when constructive abstractions arise under the strong causal abstraction framework developed by Beckers and collaborators. The key result that under this framework, intervention composition consistency forces constructive abstraction is very interesting. The follow up results about establishing necessity when intervention domain is restricted also fits the narrative of the paper well and make the results complete - we now understand that under an appropriate intervention domain, intervention composition consistency and constructive abstraction are equivalent under Beckers definition of intervention mapping.

Critical weaknesses

None

Questions for authors

It’s possible I have some misunderstanding which motivated this question and I apologise in advance. The intervention mapping omega is defined via set equality via the images under tau of low level intervened set with the high level intervened set. This doesn’t capture the changes in low-level and high-level state correspondences - for example, let X1 and X2 be two disconnected low level variables taking values in the realnumbers and they map to Y via tau_Y(x1, x2) = x1+x2. Then take a low level intervention do(X1=0). Then tau_Y(RST(V_L, I)) = \mathbb{R} and so omega should map this low level intervention to the empty high level intervention. But clearly, the high-level states are impacted by the low level intervention: say we start with unintervened low level states x1,x2 = 1,1 , the corresponding high level state is y= 2, then as we change x1 to 0, y should change to 1; this change is not captured by the high level intervention supposed by the set equality requirement. If we no longer suppose this way of mapping interventions from low level to high level, the paper’s result about intervention composition forcing constructive intervention no longer holds - so i wonder what motivates us to define an intervention mapping using set equalities in the first place since they don’t reflect state-wise high-level-low-level correspondence under interventions?

1 comment

We appreciate your positive assessment of our theoretical results and your thoughtful question about the intervention mapping.


1. Limitations of set equality and strong abstraction

Your observation about set equality is correct. Your example is exactly the real-valued sum example discussed in Section 4.1 (lines 245–248 of the original manuscript). This also motivates Section 4, where we suggest not imposing Beckers and Halpern's strong abstraction as a general requirement.

The issue of set equality and strong abstraction.

Set equality condition compares the sets of states allowed by the two interventions, rather than checking what happens to each individual state. A low-level and high-level intervention correspondence which satisfies state-wise consistency definitely will change the set of allowed states correctly, but the converse is not true. Therefore, intuitively and informally, set equality condition is necessary, but weak and coarse-grained. Take your sum example: do⁡(X1=a,X2=b)\operatorname{do}(X_1=a, X_2=b) naturally corresponds to do⁡(Y=a+b)\operatorname{do}(Y=a+b). This is a good correspondence that satisfies both set equality and interventional consistency. But do⁡(X1=a)\operatorname{do}(X_1=a) with the high-level empty intervention also satisfies set equality but not state-wise consistency. And Strong abstraction requires every such set-equality correspondence to satisfy state-wise consistency. This can exclude natural aggregation mappings such as summation.

What motivates defining an intervention mapping through set equality?

Our understanding is that Beckers and Halpern wanted to decouple the choice of intervention correspondences from the check of their effects on the models. Given a low-level SCM, a high-level SCM, and a state mapping τ\tau, their framework separates two steps:

  1. Use τ\tau to determine ωτ\omega_\tau through set equality.
  2. Check whether the resulting low-level and high-level intervention correspondences satisfy interventional consistency.

The first step uses only τ\tau and the state spaces, independently of the SCM structures. So we can uniquely determine ωτ\omega_\tau from τ\tau in the first place. And whether the outcomes agree state by state, including in your example, is checked in the second step.

Although set equality alone does not guarantee state-wise correspondence, defining ωτ\omega_\tau from τ\tau through set equality is central to Beckers and Halpern’s framework. This framework, building on Rubenstein et al. (2017), is an established framework for causal abstraction. Changing how ωτ\omega_\tau is defined would amount to change their whole definition of causal abstraction framework, so we prefer to address the issue within their framework. We agree that the definition is not perfect and exploring how to revise it is interesting future work.


How can we address the issue within this framework?

Beckers and Halpern define both τ\tau-abstraction and strong τ\tau-abstraction. The former lets us choose the allowed intervention domain and check consistency on that domain. The latter requires all high-level hard interventions and all compatible low-level interventions. So our solution is simple: use τ\tau-abstraction, which allows us to restrict the intervention domain. In Section 4, we propose the domains Jτ\mathcal J_\tau and Jτrel\mathcal J_\tau^{\mathrm{rel}}. These retain every realizable high-level intervention while excluding interventions such as do⁡(X1=a)\operatorname{do}(X_1=a) in your sum example.


2. Does our result still hold without assuming set equality?

In fact, composition preservation can still force constructivity when we change the causal abstraction framework. If we replace Beckers and Halpern's framework with Rubenstein et al.'s (2017) exact transformations, ω\omega can be specified separately; it no longer needs to be defined from τ\tau through set equality. We have added this result as Theorem 5 in Section 4.3 of the revised manuscript. Under the theorem's stated assumptions, the result can be summarized as

interventional consistency+composition preservation;⟹;constructivity.\text{interventional consistency} + \text{composition preservation} \\;\Longrightarrow\\; \text{constructivity}.

The intuition is just what we discussed above: set equality is a relatively weak requirement. In the proof, interventional consistency and composition preservation, together with the theorem's coverage and domain conditions, already imply set equality. Once we recover this connection between τ\tau and ω\omega, the rest of the proof follows the same reasoning used to establish constructivity in Theorem 3 of the original manuscript. The result also holds if we use state-wise consistency instead of distributional consistency.

Beckers, S. & Halpern, J. Y. (2019). Abstracting Causal Models. AAAI 2019, 33(01), 2678–2685.

Rubenstein, P. K. et al. (2017). Causal Consistency of Structural Equation Models UAI 2017, 151–160.

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

The authors study an important theoretical question in the definition of τ\tau-abstractions between Structural Causal Models. In particular, they focus on the relation between strong abstractions, where each high-level intervention must find a low-level implementation, and constructive abstraction, where each high-level variable must depend on a disjoint set of low-level variables. In this context, they build on Beckers and Halpern (2019), who conjectured that strong abstraction entails constructive abstraction, and Massidda et al. (2024), who proved the conjecture for linear Additive Noise Models. The authors study the non-parametric setting, showing that the conjecture does not hold under common regularity conditions, and detail a further condition, i.e., composition preservation, under which it holds. They then discuss the problem of restricting the set of allowed low-level interventions in strong abstraction and prove that, on these restricted domains, composition preservation is equivalent to constructive abstraction. Finally, they report a small synthetic experiment in which they exploit composition to regularize the learning of a nonlinear abstraction function.

Critical strengths

The authors tackle an important conjecture in causal abstraction and their discussion is rigorous and well proven. In particular, the authors first prove that if the intervention map preserves composition, then strong abstraction implies constructive abstraction for non-parametric models (Theorem 2). Furthermore, they show that if low-level interventions are allowed only on the variables relevant to the intervened high-level variables, then composition preservation is a necessary and sufficient condition for constructive abstraction (Theorem 3).

Critical weaknesses

Theorem 1 relies on cancellation. The authors claim that any edgeless model admits any arbitrary strong abstraction "despite strong regularity conditions". The claim, proven in Theorem 1, builds on a generalization of the assumptions for the linear case introduced by Massidda et al. (2024). However, their example relies on exactly one of the scenarios that the linear assumptions aimed to exclude, i.e., cancellations in the abstraction function. In the linear case, the Jacobian and the linear map coincide, and τ\tau-abstraction requires non-zero coefficients. The authors generalize this by asking the Jacobian to be non-zero somewhere, instead of non-zero always. Their counterexample works because of this property, since the abstraction function depends on some of the relevant variables only for certain configurations of the remaining ones. For instance, Fj(p,(a,a))=aF_j(p, (a, a)) = a for every parent value pp, so ∂τj/∂zr,s=0\partial \tau_j / \partial z_{r,s} = 0 whenever zj,0=zj,1z_{j,0} = z_{j,1}, even though zr,s∈Rjz_{r,s} \in R_j. This allows one to "copy" some values only when needed. While I invite the authors to explore this scenario, the claim as it stands only covers a generalization that misses the motivation for the original linear assumptions, thus overselling the "natural regularity conditions" under which constructivity might fail.

Composition preservation constrains ω\omega. The conjecture of Beckers and Halpern (2019) deals with two strongly consistent causal models and constructivity of the abstraction map as a necessary condition. The authors introduce a requirement directly on the intervention map ω\omega, which in the causal abstraction literature is directly entailed by the abstraction function τ\tau and not vice versa. The results therefore tackle the conjecture with a condition on ω\omega that is close to constructivity by design; for instance, the composition fails in Example 4.1 because the interventions share a low-level variable.

Repair of Massidda et al. (2024). The authors claim to repair results from Massidda et al. (2024). In particular, the authors claim that Examples 1, 3, 5 and Theorem 3 from that paper fail due to considering all possible low-level interventions to have a corresponding high-level intervention. However, in the proof of Lemma 1 from Massidda et al. (2024), they already prove that any low-level implementation of a high-level intervention must cover all and only its relevant set. The rest of the paper then only considers interventions of this form, ignoring the others that do not consistently correspond to a high-level intervention. The authors of the submission introduce a similar condition for non-parametric causal models, which coincides with the one proposed by Massidda et al. (2024) in the linear case (Lemma F.1). Therefore, while the authors correctly spotted an underrepresented claim in the main body of Massidda et al. (2024), hence clarifying an interesting point for the causal abstraction community, the claim of "repairing" results of the previous paper might be misleading.

Minimal learning example. The learning example is preliminary and not sufficiently detailed with respect to its connection to the theory. It targets the same family of functions used in Theorem 1, restricted to be constructive, and the composition constraint relies on the same interventions setting pairs to equal values. Therefore, the limitations of the example I reported earlier also hold for this learning example. Moreover, the algorithm assumes access to abstract samples. It should be clearly stated that the problem is therefore fully supervised.

Presentation. Despite my familiarity with Causal Abstraction, I found the paper hard to follow. In particular, the restricted domains JτJ_\tau and JτrelJ_\tau^{rel} are defined through the targets of ωτ\omega_\tau, and the main text never discusses what they correspond to in simple cases. For instance, the fact that in the linear case they reduce to interventions on whole relevant sets (Lemma F.1) is only shown in the appendix, while it is fundamental to compare with previous work. Furthermore, several terms are used without an explicit definition, such as state cylinders and full-domain consistency. Finally, the authors defer most of the argument to the appendices, i.e., the construction of Theorem 1 (Appendix A), the closure under composition (Appendix C.1), and the scope of the repairs (Appendix F). Therefore, the main text alone does not allow the reader to assess either the main counterexample or the repair claims.

Limited discussion. The conclusions of the paper are limited in scope, as they discuss neither the impact of proving the relation between strong and constructive abstraction, nor the scenarios under which the authors expect their assumptions to be realistic.

Questions for authors
  • Have you considered the scenario where the abstraction function never cancels low-level variables, strictly generalizing the linear scenario?
  • The paper often refers to state cylinders, while Definition 2.2 only introduces the notation Rst\mathrm{Rst}. Could you provide an explicit definition?
  • Theorems 2 and 3 do not seem to use the interventional consistency between the low-level and the high-level models, and only depend on τ\tau and ωτ\omega_\tau. Is this the case? If so, how do they relate to the conjecture of Beckers and Halpern (2019), which concerns causal models?
7 comments

Thank you for your careful review, detailed comments and suggestions, especially related to discussing the connection between our submission and Massidda et al. (2024). Below, we respond to each weakness and question.


Weakness 1 and Question 1

Theorem 1 relies on cancellation

Yes, your observation about Theorem 1 is correct: the construction relies on cancellation. Since no-cancellation is an important assumption in Massidda et al. (2024) and it is another way to imply constructivity, we will remind readers of this and add a discussion in the revised manuscript. Below is part of the exploration we make and the details are in the revised manuscript.

In fact, sufficiently strong no-cancellation assumptions, together with interventional consistency, can imply constructivity, assuming acyclicity and full high-level intervention coverage. Below is an intuitive example of proof (this is similar to the proof of Lemma1 Massidda et al. (2024)).

Consider

Y1=X1+X2,Y2=X2+X3.Y_1=X_1+X_2,\qquad Y_2=X_2+X_3.

This mapping is not constructive. We show why sufficiently strong no-cancellation assumptions prevent interventional consistency in this case.

By acyclicity, we can swap the labels so that Y1Y_1 is not an ancestor of Y2Y_2. Choose one allowed low-level implementation ii of (Y1←a)(Y_1\leftarrow a). Since τ1=X1+X2\tau_1=X_1+X_2, this implementation must fix both relevant inputs X1X_1. X2X_2. Interventional consistency for ii and the empty intervention therefore gives

Y2=τ2(MLi(e))=τ2(ML(e)).Y_2=\tau_2(M_L^i(e))=\tau_2(M_L(e)).

Fixing X2X_2 makes the left-hand side independent of X2X_2 's private noise E2E_2, contradicting the dependence required by no-cancellation on the right. Here, "private" means that E2E_2 enters only the equation for X2X_2 and can vary while the other noises are held fixed.

This argument illustrates two requirements that together suffice for constructivity:

  1. Condition E. For each high-level variable YjY_j, there is some value aa and an allowed low-level intervention that fixes all variables in RjR_j and maps to do(Yj=a)do(Y_j=a) alone. This holds for linear mappings under high-level intervention coverage, but may fail for nonlinear mappings.
  2. No-cancellation for τ∘ML\tau\circ M_L. Each relevant variable's private noise must retain an influence on the corresponding reduced-form output.

Everywhere-nonzero derivative condition mentioned by reviewer is a stronger version of E. The second requirement addresses cancellation after composing τ\tau with the low-level model, as discussed by Massidda et al. (2024, Section 2).

Our existing examples illustrate why both requirements matter. In Theorem 1, ML(e)=eM_L(e)=e, so relevant private-noise dependence survives, but E fails in the non-constructive cases. Conversely, in Example 4.1, E holds, but X2=E2X_2=E_2 and X3=E3−X2X_3=E_3-X_2 give Y2=X2+X3=E3Y_2=X_2+X_3=E_3: the private noise cancels. That example remains non-constructive and interventionally consistent on its restricted domain.


Overselling the “natural regularity conditions” under which constructivity might fail.

Sufficiently strong no-cancellation assumptions can imply constructivity, but they need not be “minor technical conditions” in the general nonlinear setting.

No-cancellation assumptions have a natural motivation in linear causal discovery. For linear mappings, the first requirement holds automatically. The second type of cancellation requires exact cancellation of coefficients; under a freely varying linear parameterization, zeros of a nontrivial total-effect polynomial form a Lebesgue-null set. This parallels the genericity argument used to motivate faithfulness in linear models (Meek, 1995).

However, in the general setting considered here, these two requirements can impose substantive restrictions.

First, consider the simple gain model τ(x,g,b)=(x,gx+b)\tau(x,g,b)=(x,gx+b), with outputs A=XA=X and Y=GX+BY=GX+B. Setting G=0G=0 and B=yB=y implements do(Y=y)do(Y=y) while leaving AA free, and is allowed on our restricted domains. However, fixing all relevant inputs of YY, namely X,G,B\\{X,G,B\\}, also fixes AA. Thus no implementation fixing all these inputs can target YY alone: E fails. With independent low-level roots, τ∘ML\tau\circ M_L still satisfies the second requirement.

Second, the private-noise formulation above requires non-degenerate local noise, which need not exist when a system contains deterministic mechanisms. This is particularly relevant to mechanistic interpretability, an important application of causal abstraction: the low-level variables are often neural activations, computed deterministically from preceding layers without an additional private noise input at each neuron (Geiger et al., 2025, Section 2).

This is similar as faithfulness, which are natrual in linear causal discovery but be changlenged more often in nonlinear (Marx et al. (2021)) and deterministic (Garrabrant et al., 2024, Li et al. (2024)) setting.

Weakness 2

The results therefore tackle the conjecture with a condition on ω\omega that is close to constructivity by design

Constructivity is a property of the mapping τ\tau: no low-level variable XsX_s may influence two distinct high-level outputs YiY_i and YjY_j through τ\tau.

Composition preservation is an algebraic property of the intervention mapping ωτ\omega_\tau:

ωτ(i∘k)=ωτ(i)∘ωτ(k).\omega_\tau(i\circ k)=\omega_\tau(i)\circ\omega_\tau(k).

The two conditions therefore concern different aspects of abstraction, and proving their equivalence is nontrivial. You are correct that ωτ\omega_\tau is determined by τ\tau through set equality. This set equality provides the bridge for proving both directions: that composition preservation of ωτ\omega_\tau implies constructivity of τ\tau, and that constructivity of τ\tau implies the homomorphism property of ωτ\omega_\tau.

In practice, this characterization can motivate approaches to testing constructivity when the form of ω\omega is unknown but interventional data are available. In the other direction, it shows that, in a constructive abstraction, arbitrary finite sequences of allowed interventions preserve interventional consistency between the intervened models under the same τ\tau and ωτ\omega_\tau.

for instance, the composition fails in Example 4.1 because the interventions share a low-level variable.

Yes. We included this example to illustrate Theorem 3, which states that composition preservation is equivalent to constructivity. Its purpose is to show how non-constructivity can cause composition preservation to fail. The fact that the example is intuitive does not make the general theoretical result straightforward or trivial.

Weakness 3

Intervention domain in Massidda et al. (2024).

We agree that our "repair" does not affect the nontrivial and valuable contributions of Massidda et al. (2024). Their main text specifies a full intervention domain, while their proof works with interventions on complete relevant sets. We will change the term "repair" to "clarification" to make clear that we are highlighting an underrepresented claim without changing the substance of their results.

Weakness 4

The learning example is preliminary and not sufficiently detailed with respect to its connection to the theory.

Yes, the experiment is preliminary. Our paper is mainly theoretical and the experiment part is just to show how comp can motivate a loss term. We never claim that we proposed a causal discovery method for identifying causal abstraction in general cases. And in the revised manuscript, we will explain more clearly how the loss is motivated by intervention composition preservation.

How does the loss terms we proposed relate to the comp condition?

Take Output Comp (Eq. 7) as an example:

LoutComp=13B∑r=1B∑j≠kr[τθ,j(x′(r))−τθ,j(x~(r))]2.\mathcal L_{\mathrm{out}}^{\mathrm{Comp}} = \frac{1}{3B}\sum_{r=1}^{B}\sum_{j\ne k_r} \left[ \tau_{\theta,j}(x'^{(r)}) - \tau_{\theta,j}(\tilde x^{(r)}) \right]^2.

Our theoretical composition condition is

ωτ(i∘k)=ωτ(i)∘ωτ(k).\omega_\tau(i\circ k)=\omega_\tau(i)\circ\omega_\tau(k).

For sample rr, take k=(VL←x~(r))k=(V_L\leftarrow\tilde x^{(r)}), which fixes all low-level variables, and i=((akr,bkr)←(cr,cr))i=((a_{k_r},b_{k_r})\leftarrow(c_r,c_r)), which overwrites one input pair. Their composition fixes the low-level state to x′(r)x'^{(r)}.

The left-hand side therefore fixes all high-level variables to τθ(x′(r))\tau_\theta(x'^{(r)}). The right-hand side first fixes them to τθ(x~(r))\tau_\theta(\tilde x^{(r)}), then overwrites only YkrY_{k_r} with crc_r. Thus, Comp requires

τθ,j(x′(r))=τθ,j(x~(r)),j≠kr.\tau_{\theta,j}(x'^{(r)})=\tau_{\theta,j}(\tilde x^{(r)}), \qquad j\ne k_r.

Therefore we penalize τθ,j(x′(r))−τθ,j(x~(r))\tau_{\theta,j}(x'^{(r)})-\tau_{\theta,j}(\tilde x^{(r)}) in our loss.

We also agree that the setting is fully supervised and we will state this explicitly.


Weakness 5

We agree that the presentation can be clearer and thanks for helping us to improve readabiltiy. In the revised manuscript, we will explicitly introduce the terms "state cylinder" and "full-domain consistency", move the block-linear interpretation of the restricted domain from Lemma F.1 into the main text, and improve the readability of our paper.

For the restricted domains JτJ_{\tau } and JτrelJ_{\tau }^{rel} and the construction of Theorem 1, the original main text already illustrates these with a sum example in Figure 3 and outlines the construction of Theorem 1 in Section 3.


Weakness 6

Thank you for pointing this out. With one more page scope allowed in the rebuttal period, we will expand the discussion to explain both the implications and the scope of our assumptions in the revised manuscript.

In short, our theoretical results, from one side, can test the constructivity, for example, if we find that a composition of preservation is not consistent, then we can quickly know constructivity fails. And from the other side, our results give supporting evidence to do multiple times of interventions in constructivity scenarios.


Question 2

Yes, we will explicitly define "state cylinder" when introducing Rst⁡\operatorname{Rst} in the revised manuscript. A state cylinder is the set of all states in which the intervened variables have their assigned values, while all other coordinates are unrestricted.


Question 3

Yes, your observation is correct. The proofs of Theorems 2 and 3 do not use state-wise interventional consistency between the two models. Instead, they use the set-equality condition

τ(Rst⁡(VL,i))=Rst⁡(VH,ωτ(i))\tau(\operatorname{Rst}(V_L,i))=\operatorname{Rst}(V_H,\omega_\tau(i))

as a bridge between the properties of τ\tau and those of ωτ\omega_\tau.

How does this relate to Beckers and Halpern's conjecture, which concerns causal models and state-wise consistency?

State-wise consistency relates the outcomes of the low-level and high-level models. The set-equality condition can also be viewed as a form of consistency between the two levels, but it is weaker and more coarse-grained: it concerns the state constraints imposed by interventions and does not depend on the models' specific mechanisms.

Intuitively, a correspondence between low-level and high-level interventions that satisfies state-wise consistency also constrains the corresponding ranges of possible outcomes. The converse, however, is not guaranteed.

In the revised manuscript, we add Section 4.3, where we do not impose set equality. Instead, we adopt the intervention mapping and distributional consistency of Rubenstein et al.'s framework. Composition preservation still implies constructivity under the assumptions stated there. This subsection also clarifies the relationship between set equality and state-wise or distributional consistency.

Thus, our results do concern causal models, but the original proofs rely on a weaker, mechanism-independent form of consistency between the two levels.


Beckers, S. & Halpern, J. Y. (2019). Abstracting Causal Models. AAAI 2019, 33(01), 2678–2685.

Garrabrant, S., et al. (2024). Factored Space Models: Towards Causality Between Levels of Abstraction. arXiv preprint arXiv:2412.02579.

Geiger, A., et al. (2025). Causal Abstraction: A Theoretical Foundation for Mechanistic Interpretability. JMLR, 26(83):1–64.

Li, L., et al. (2024). On Causal Discovery in the Presence of Deterministic Relations. NeurIPS 2024.

Marx, A., Gretton, A. & Mooij, J. M. (2021). A Weaker Faithfulness Assumption Based on Triple Interactions. UAI 2021.

Massidda, R., Magliacane, S. & Bacciu, D. (2024). Learning Causal Abstractions of Linear Structural Causal Models. UAI 2024.

Meek, C. (1995). Strong Completeness and Faithfulness in Bayesian Networks. UAI 1995.

Rubenstein, P. K. et al. (2017). Causal Consistency of Structural Equation Models. UAI 2017, 151–160.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of revisions

We thank the reviewers for their constructive feedback. The main revisions, highlighted in blue, are:

  1. Presentation (ybtL: W5, Q2; CK2p: W2, Q1). We clarify key definitions and distinguish mapping constructivity from constructive τ\tau-abstraction in Section 2.

  2. Interventional consistency (ybtL: Q3; A4s3: question on set equality). New Section 4.3 establishes a constructivity result without assuming set equality, under the stated conditions.

  3. No-cancellation (ybtL: W1, Q1). We compare no-cancellation with composition preservation and examine its scope in nonlinear settings, with details in the appendix.

  4. Intervention domains (ybtL: W3, W5). We add the block-linear interpretation to the main text and replace “repair” with “clarification” for Massidda et al. (2024).

  5. Learning example (ybtL: W4; CK2p: W3). We explain how Comp motivates the loss and explicitly identify the setting as fully supervised.

  6. Discussion (ybtL: W6; CK2p: W1, Q3). We expand the conclusion to discuss the implications and scope of our results and assumptions.

AuthorsOpenReview ↗

Summary of the second revision

Following Reviewer ybtL’s follow-up, we make three changes:

  1. In the introduction, we clarify that our goal is to characterize constructivity of a given τ\tau through its induced intervention map ωτ\omega_\tau, rather than to give direct guidance on how to build a constructive abstraction map.
  2. We simplify the derivation of the composition loss in Section 5 and clarify the practical limitation of our “omega-constraining” strategy. Computing the loss requires knowing how some low-level interventions correspond to high-level interventions.
  3. In Figure 5(b), we replace sampled composition MSE with a test-set Jacobian measure of each output’s sensitivity to inputs outside its assigned pair.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.