On the Identifiability of Causal Abstractions

Xiusi Li, Sékou-Oumar Kaba, Siamak Ravanbakhsh

Abstract

Causal representation learning (CRL) enhances machine learning models’ robustness and generalizability by learning structural causal models associated with data-generating processes. We focus on a family of CRL methods that uses contrastive data pairs in the observable space, generated before and after a random, unknown intervention, to identify the latent causal model. While previous work showed identifiability is possible if all latent variables can be intervened on individually, this is often unrealistic. In this work, we instead assume interventions on arbitrary subsets of latent variables. We introduce a theoretical framework that calculates the degree to which we can identify a causal model, given a set of possible interventions, up to an abstraction that describes the system at a higher level of granularity.

Introduction

Causal Representation Learning (CRL) is an emerging field that seeks to bridge the gap between high-dimensional observations—such as pixels or sensor readings—and the underlying structural causal models (SCMs) that govern them. By uncovering these hidden variables and their causal relationships, AI systems can achieve better out-of-distribution generalization and more robust reasoning. However, a significant bottleneck in current CRL theory is the reliance on "atomic interventions," where each latent variable must be capable of being isolated and manipulated independently.

Presented at AISTATS 2025 by researchers from Mila and McGill University, this paper challenges that requirement. The authors demonstrate that even with "messy," overlapping interventions on subsets of variables, we can still recover a meaningful, interventionally consistent version of the truth, known as a causal abstraction.

The Problem: The Atomic Intervention Requirement

Most existing proofs of identifiability in CRL assume that for every hidden variable, there exists an intervention that changes only that variable. In complex real-world systems—like biological pathways or economic markets—this is rarely the case. Intervening on one gene often triggers a cascade of changes in others, and policy shifts rarely affect a single economic indicator in isolation.

A ground truth causal graph with five latent variables, representing the underlying structure to be recovered.

As shown in the figure above, a system might have multiple interconnected variables. If we can only perform interventions on groups of these variables (e.g., intervening on nodes 3 and 4 simultaneously), traditional CRL methods struggle to disentangle the individual components. This paper asks: what can we still guarantee about the causal model we recover when atomic interventions are unavailable?

Methodology: Learning Through Abstraction

To address this, the authors frame the problem using Causal Abstractions. Instead of attempting to recover the "microscopic" ground truth perfectly, they aim to recover a "macroscopic" model where variables are grouped into identifiable blocks.

Key components of their methodology include:

  • Counterfactual Pairs: The model learns from data pairs $(x, \tilde{x})$ representing the system state before and after a random, unknown intervention.
  • SCM Homomorphisms: Using category theory, the authors define how a complex SCM can be mapped to a simpler, abstracted version while maintaining causal consistency. This mapping ensures that the "coarse" model still responds to interventions in a way that reflects the underlying truth.
  • Graphical Criteria: The framework utilizes the non-descendants of intervention targets. If an intervention occurs, the variables that remain unchanged provide a critical signal for defining the boundaries of identifiable variable blocks.
The transformation of a complex latent causal model into a coarser, identifiable abstraction (quotient graph).

Key Insights: The Spectrum of Identifiability

The paper's most significant contribution is the realization that identifiability is not a binary "all or nothing" property. Instead, it is a spectrum determined by the "overlap" of available interventions.

  1. The Power of Invariance: You do not necessarily need to intervene on a variable to identify it. By observing which variables remain invariant across different subset interventions, the model can deduce the boundaries of latent factors.
  2. The $\sigma$-algebra Partition: The authors prove that identifiable blocks of variables correspond exactly to a partition generated by the $\sigma$-algebra of the non-descendant sets of the intervention targets. This provides a clear "graphical recipe" for researchers to determine exactly what level of detail is learnable in any given system.
  3. Disentanglement without the Full Graph: Even if the entire causal graph cannot be recovered, specific latent variables can still be "disentangled" if they are uniquely isolated by the intersections of the available intervention sets.

Results and Validation

The theoretical findings are supported by two core theorems:

  • Theorem 3.1: Proves that any latent causal model is identifiable up to an SCM abstraction, where the recovered causal graph is a "quotient graph" of the true graph.
  • Theorem 3.2: Provides the specific conditions under which individual latent variables can be recovered even when the full structural model cannot be.

To validate these theorems, the authors conducted experiments using synthetic Linear Gaussian Models. Latent variables were transformed into observations via a random rotation matrix (an invertible mixing function). In these tests, the learned encoder successfully recovered the ground truth mixing matrix up to a block-diagonal structure, with the size and position of the blocks matching the theoretical predictions exactly.

Conclusion and Future Work

This work significantly relaxes the assumptions required for Causal Representation Learning, making it more applicable to real-world data where precise interventions are impossible. While the current proof relies on counterfactual pairs and assumes an invertible mixing function (diffeomorphism), it opens the door for future research into observational settings and non-linear, high-dimensional data like video. By shifting the goal from "perfect recovery" to "identifiable abstraction," the authors provide a more realistic and mathematically grounded path forward for causal AI.

Summary updated

Abstract

Causal representation learning (CRL) enhances machine learning models’ robustness and generalizability by learning structural causal models associated with data-generating processes. We focus on a family of CRL methods that uses contrastive data pairs in the observable space, generated before and after a random, unknown intervention, to identify the latent causal model. (Brehmer et al., 2022) showed that this is indeed possible, given that all latent variables can be intervened on individually. However, this is a highly restrictive assumption in many systems. In this work, we instead assume interventions on arbitrary subsets of latent variables, which is more realistic. We introduce a theoretical framework that calculates the degree to which we can identify a causal model, given a set of possible interventions, up to an abstraction that describes the system at a higher level of granularity.
BibTeX
@inproceedings{li2025identifiability,
  author = {Xiusi Li and Sékou-Oumar Kaba and Siamak Ravanbakhsh},
  title = {{On the Identifiability of Causal Abstractions}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2025},
  url = {https://openreview.net/forum?id=JkB5ZknFsX},
}
Figure