Identifiability Guarantees for Causal Disentanglement from Purely Observational Data

Causal disentanglement aims to learn about latent causal factors behind data, holding the promise to augment existing representation learning methods in terms of interpretability and extrapolation. Recent advances establish identifiability results assuming that interventions on (single) latent factors are available; however, it remains debatable whether such assumptions are reasonable due to the inherent nature of intervening on latent variables. Accordingly, we reconsider the fundamentals and ask what can be learned using just observational data. We provide a precise characterization of latent factors that can be identified in nonlinear causal models with additive Gaussian noise and linear mixing, without any interventions or graphical restrictions. In particular, we show that the causal variables can be identified up to a layer-wise transformation and that further disentanglement is not possible. We transform these theoretical results into a practical algorithm consisting of solving a quadratic program over the score estimation of the observed data. We provide simulation results to support our theoretical guarantees and demonstrate that our algorithm can derive meaningful causal representations from purely observational data.

Paper

Similar papers

Peer review

Reviewer gAcJ6/10 · confidence 4/52024-06-26

Summary

The authors propose a method to identify causal and exogenous variables in Gaussian Additive Noise Models from purely observational data.

Strengths

The paper proposes a novel approach (the setting might have been largely considered elsewhere; see my comments in **Weaknesses** and **Questions**). - the proofs (I checked a significant part of the appendix, but not everything) seem to be sound - the identifiability notion up to upstream layers is an interesting concept (needs to be better discussed what it means and how it compares to other notions) I chose a conservative score; however, if the authors address my concerns (mostly about making statements more precise) in a satisfying manner, I will increase my score.

Weaknesses

## Major points - Your statement _our work is the first to establish identifiability guarantees in the purely observational setting without imposing any structural assumptions over the mixing function_ seems to be missing some nuance: - assuming an ANM is a structural assumption (and, at least from a theoretical perspective, a rather strong one) - you cite [32], which shows that nonlinear ICA can also uncover the causal graph. As I consider nonlinear ICA methods to be purely observational, and as [32] does not assume additivity, this seems to contradict your claim (as, in that case, you would get the DAG and the exogenous variables). Though in [32], there is no level corresponding to $H$ -- if this is the novely, then please be more precise. - optimizing the Jacobian is computationally very expensive; I haven't found any discussion on this aspect - a more detailed discussion on delineating the contribution would be highly suggested (from both score-based methods and [32]) - I also encourage the authors to address the (limitations of) identifiability up to upstream layers; though, as I wrote above, find the concept interesting, my understanding is that this is a very limited identifiability notion ## Minor points - line 93: "principal" -> "principle" - use different $C, P_\pi$ for Defs 1 and 2 - the definition of, e.g., $\beta$ in Lem 1 comes a bit late; this makes the reader wonder what it stands for - resolve each abbreviation and symbol in each Table and Figure caption, even if you defined those quantities in the main text

Questions

- What exactly do you mean by asymmetries? I only found the related line 46, and that does not seem to be any novel key insight, so I am assuming I am missing your point. - What do you exactly mean by causal disentanglement? - Why do you consider Gaussian exogenous variables when that causes the (unsurprising) impossibility result in 3.3 - Can you use insights from score-matching-based methods to improve your identifiability results (those methods can easier identify leaf nodes, whereas your method works better for root nodes; I reckon you draw some connections, but I could not find this specific aspect) - In Fig. 4, what is the MAC for SSM and STEIN? Am I interpreting it correctly that it can be everything? - How does identifiability up to layers (particularly, distinguishing leaf and non-leaf variables) relate to block-identifiability (if it does relate), or, particularly, to content and style latent variables (cf. [1]) ## References - [1] Von Kügelgen, J., Sharma, Y., Gresele, L., Brendel, W., Schölkopf, B., Besserve, M., & Locatello, F. (2021). Self-supervised learning with data augmentations provably isolates content from style. Advances in neural information processing systems, 34, 16451-16467.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors addressed broader impact. I asked for clarifications regarding computational and novelty aspects.

Authorsrebuttal2024-08-08

Response to comment

Thank you for the discussion and for updating the scores! Thanks for the suggestion! We will revise the manuscript to clarify the concepts regarding (1) causal representation learning and causal disentanglement, and (2) structural restrictions and graphical restrictions. Our usage and initial thoughts aim to align with recent prior works in this area (e.g., the usage of causal disentanglement in [1], and structural restrictions in [2]). Nonetheless, we will clarify these concepts to improve readability. In addition, we would like to comment on why we believe that while ANM poses certain assumptions, it can still be a useful model in many scenarios: **The Additive Noise Model Assumption** When compared to alternative assumptions (e.g., independent components, which are a special case of ANM) in identifiable representation learning, we believe that ANM is relatively less restrictive, as it allows dependencies between the latent components. These dependencies are important, in our opinion, as the latent factors should be related to each other. Additionally, in the causal structure learning setting where all causal variables are observed, ANMs are frequently used because (1) their theoretical properties are well understood, (2) efficient methods exist for learning them, and (3) they can fit non-parametric relationships, making them flexible for modeling many real-world causal systems, such as gene regulatory networks [3]. Therefore we believe that it would be helpful to devise understandings and methods to extend these models to the causal representation learning setting, where we can handle perceptual and non-tabular observations. --- References: [1] Causal Machine Learning: A Survey and Open Problems \ [2] Linear causal disentanglement via interventions \ [3] Estimation of genetic networks and functional structures between genes by using Bayesian networks and nonparametric regression

Reviewer gAcJ2024-08-08

Score re-evaluation

Thanks for the detailed response. I went through your earlier response again, and in this light, I modified my score to 6.

Authorsrebuttal2024-08-08

Response

Thank you again for the discussion! We really appreciate the suggestions.

Reviewer 2B2i4/10 · confidence 4/52024-07-12

Summary

This paper concerns itself with the problem of causal disentanglement from purely observational data. The setting is that data X is generated as X = H.Z where H is a linear matrix and Z are latent variables that follow a nonlinear additive Gaussian noise model. In this setting, it is shown that the latent variables can be identified upto upstream layers (defn 1), when given access to infinite data. The idea is to compute the score functions (gradient of log-likelihood) and use their structural properties to recover the layers. This is then converted into an algorithm, which computes the desired score functions, using standard score-estimation methods and an application of quadratic programming to identify the model. The technique is validated numerically via simulated experiments. This setting is similar to [1] however interventional data is assumed to not be available, however, the identifiability results seem weaker as expected. However, the claims made in the paper seem too strong (e.g. this is not the first paper to consider the purely observational setting, see [2]) and the proof techniques also seem to have appeared in prior works such as [3] (see weaknesses). ### References: - [1] Score-based causal representation learning with interventions. B. Varici, E. Acarturk, K. Shanmugam, A. Kumar, A. Tajer. - [2] Identifiability of deep generative models without auxiliary information. B Kivva, G Rajendran, P Ravikumar, B Aragam - [3] Score matching enables causal discovery of nonlinear additive noise models. P Rolland, V Cevher, M Kleindessner, C Russell, D Janzing, B Schölkopf, F Locatello

Strengths

- Compared to a flurry of recent works that assume existence of interventional data to recover latent variables in data, this work studies purely observational data, which is a more useful setting. - The notion of identifibility upto layer-wise transformations is novel to the best of my knowledge and could potentially be useful elsewhere.

Weaknesses

- L184/L284 claims there are no structural results on the mixing function, however assumption 1 assumes that the mixing function is linear, which is a pretty big restriction. This weakens the claim of this paper. - L81-82 says "our work is the first to establish identifiability guarantees in the purely observational setting without imposing any structural assumptions over the mixing function". However, the prior work [2] also considers the purely observational setting and the assumption they make on the mixing function is piecewise-linearity, which is much weaker than the linearity assumption made here. That makes this claim not valid, could the authors clarify this point? - This work seems to be an extension of [3] where an additional linearity is added. The lemmas and proofs also seem similar. Could the authors describe the main differences in the assumptions and/or additional difficulties in the proof? - Experiments are on synthetic data. While they weakly validate the theoretical results, an application to real-life would be needed to see if these ideas extend to practice.

Questions

Some questions were raised above. - Identifiability up to layers seems a bit hard to grasp. While theoretically true, what does it mean for the practitioner? - Typo in L26: extend -> extent

Rating

4

Confidence

4

Soundness

2

Presentation

2

Contribution

1

Limitations

Limitations have been discussed.

Reviewer 2B2i2024-08-11

Score update

Thank you for your response. I understand the author’s clarifications on the differences between the works that I have cited, but my concern is more with the framing of the contributions and it still seems like overselling of the work even if there are clear technical differences from prior works. While there is some intuition behind identifiability upto layers that the authors give, I still find it hard to understand why it would be useful either theoretically or in practice. Thanks for the other technical clarifications as well, I have increased my score.

Authorsrebuttal2024-08-12

Response to further points

Thank you for the response and the further comments. We are glad that we are able to make the technical clarifications. Regarding the remaining concerns about the framing of the contributions, we really appreciate the comments and will try to clarify it via the following changes: 1. Causal disentanglement in the purely observational setting: in the related work section on causal disentanglement (line 81-82), we stated that “our work is the first to establish identifiability guarantees in the purely observational setting without imposing any structural assumptions over the mixing function”. We will modify this claim to “our work establishes identifiability guarantees of causal disentanglement in the purely observational setting, without imposing any structural assumptions over the mixing function”. In addition, we will add a discussion paragraph to clarify that identifiability of latent factors in the purely observational setting has been considered outside of causal disentanglement, such as in the work [2] mentioned by the reviewer. 2. Score-based approaches in causal discovery vs. causal disentanglement: in the related work section on “score matching in causal discovery” (line 91-92), we commented that “extending these ideas to causal disentanglement is difficult, since we do not observe the latent factors and can only estimate the log-likelihood of the observed variables”. We will add in the technical differences of our proof versus the original proof in the causal discovery setting. In particular, we will add a pointer to section 3.2 (which is where our main lemmas and theorems sit), in which we will incorporate our technical clarification in the earlier response: “The proofs in [3] utilize variance properties on the diagonal elements of the Jacobian over the score of the causal variables to derive a topological ordering. While we utilize this result, it is not sufficient given we don't have access to the specified causal variables. In Lemma 1, we prove that we can only estimate this desired Jacobian up to an unknown quadratic parameterized by the matrix , where …” We hope that these will be sufficient to clarify our contributions and avoid overselling it. Thank you for the other comments as well. In response to it, we would like to give our perspectives regarding why we believe identifiability up to layers might still be useful albeit its limitations: in the emerging field of causal disentanglement, full identifiability of latent causal model is not possible without additional assumptions, which is why many works proposed to consider, e.g., structural or sparsity restrictions on the mixing function, or access to single-node interventions. However, as we tried to illustrate in lines 39-44, these assumptions might be limiting and impractical in many settings, which is why we choose to step back and study what can be identified without interventions or structural restrictions. We choose to study nonlinear additive noise models, as they inherit the nice theoretical properties in the causal discovery setting and allow modeling of non-parametric causal mechanisms. In this case, we attain a full theoretical understanding of what can learned, by showing partial identifiability of up to causal layers, which cannot be improved without additional assumptions. Practically, this would mean that more upstream variables in a hierarchical causal structural can be disentangled easier. For example, in the context-style model in [4], our results show that the context variable can be identified up to themselves. --- References: [4] Self-supervised learning with data augmentations provably isolates content from style.

Reviewer 1YMg6/10 · confidence 3/52024-07-13

Summary

This paper studies the identifiability issue of causal disentanglement from observational data, within the setting of nonlinear causal model with additive Gaussian noise and linear mixing. An interesting result is that the causal variables can be identifiable at most up to a layer-wise transformation, based on a recent score-based causal discovery method. A practical algorithm is then proposed.

Strengths

- The paper reconsiders the fundamentals of an import problem, which is very meaningful. - The theory, in particular the concept of layer-wise identifiability, in new and motivating. - The proposed method is practical. - The paper is well-written and it is enjoyable to have a read.

Weaknesses

A key motivation of the paper relies on the reasonability of assuming interventions on latent factors. Then this paper assumes non-linear function with Gaussian additive noise and linear mixing, but these assumptions cannot be tested from observations, either. In other words, these assumptions are, in my view, alternative assumptions on the same problem, but do not relax the previous assumptions. Besides, in this setting, the dimension of latent factors is assumed to be known, which may be a limitation.

Questions

Overall, the paper studies a very interesting and meaningful problem. Within the considered setting, quite some interesting results are obtained. My questions are: 1. As mentioned in the weakness part, this paper assumes non-linear function with Gaussian additive noise and linear mixing, but these assumptions cannot be tested from observations, either. In other words, these assumptions are, in my view, alternative assumptions on the same problem, but do not relax the previous assumptions. Besides, in this setting, the dimension of latent factors is assumed to be known, which may be a limitation. Can you clarify on this? 2. the key part of identifiability theory seems to rely on the recent technique of score-based causal discovery method. In your setting, the latent factors themselves are identifiable before the linear mixing. I am wondering if other forms identifiable SCMs for the latent variables can also be identified? 3. Regarding Proposition 1.: while a counter example is sufficient for this theorem, I'd like to ask if this failure case can be avoided by putting additional assumptions? e.g., what if we put a technical condition that $a_1b_1\sigma_1^2+a_2b_2\sigma_2^2\neq 0$ or that $a_1, b_1, a_2, b_2>0$.? 4. typo: line 26: extend -> extent

Rating

6

Confidence

3

Soundness

3

Presentation

4

Contribution

3

Limitations

See above.

Reviewer zdJm6/10 · confidence 3/52024-07-17

Summary

This paper investigates causal disentanglement, learning latent causal factors from observational data without interventions. It identifies latent factors in nonlinear causal models with additive Gaussian noise and linear mixing, showing that causal variables can be identified up to a layer-wise transformation. The authors propose an algorithm based on quadratic programming over score estimation and validate it with simulations, demonstrating meaningful causal representations from observational data.

Strengths

1. The paper presents a method to identify causal variables from observational data without interventions. 2. It also provides the theoretical analysis, and demonstrates that latent variables can be identified up to a layer-wise transformation consistent with the underlying causal ordering, with no further disentanglement possible.

Weaknesses

1. The method proposed in this paper relaxes some assumptions, making it potentially more applicable to real-world scenarios. It would be better to provide the performance of applying the proposed method to real-world data. 2. In line 475, the denominator of the first term after the third equals sign in the equation should be differentiated with respect to $z_l$ instead of $k_l$? .

Questions

See the weaknesses above.

Rating

6

Confidence

3

Soundness

2

Presentation

2

Contribution

3

Limitations

Not applicable.

Authorsrebuttal2024-08-08

Response

Thank you again for the suggestion and the discussion!

Reviewer zdJm2024-08-09

Thank you for the clarifications. I’m happy to increase my evaluation score.

Authorsrebuttal2024-08-09

Thank you for the discussion and for updating the score! We are grateful to the suggestions.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC