[Paper Review] On the Identifiability of the Post-Nonlinear Causal Model

Kun Zhang, Aapo Hyvärinen|arXiv (Cornell University)|May 9, 2012

Bayesian Modeling and Causal Inference15 references244 citations

TL;DR

This paper establishes the identifiability of the post-nonlinear (PNL) causal model in the two-variable case under general conditions, proving that the causal direction can be uniquely determined from observational data when the model's nonlinearities and noise structures satisfy certain regularity conditions. It further proposes a scalable method to recover causal structures in multi-variable settings by testing conditional independence between disturbances and direct causes across Markov equivalence classes, avoiding exhaustive search over all possible causal graphs.

ABSTRACT

By taking into account the nonlinear effect of the cause, the inner noise effect, and the measurement distortion effect in the observed variables, the post-nonlinear (PNL) causal model has demonstrated its excellent performance in distinguishing the cause from effect. However, its identifiability has not been properly addressed, and how to apply it in the case of more than two variables is also a problem. In this paper, we conduct a systematic investigation on its identifiability in the two-variable case. We show that this model is identifiable in most cases; by enumerating all possible situations in which the model is not identifiable, we provide sufficient conditions for its identifiability. Simulations are given to support the theoretical results. Moreover, in the case of more than two variables, we show that the whole causal structure can be found by applying the PNL causal model to each structure in the Markov equivalent class and testing if the disturbance is independent of the direct causes for each variable. In this way the exhaustive search over all possible causal structures is avoided.

Motivation & Objective

• To establish theoretical conditions under which the post-nonlinear (PNL) causal model is identifiable in the two-variable case.
• To resolve the long-standing open problem of whether the PNL model can uniquely determine causal direction from observational data.
• To extend the PNL framework to multi-variable causal discovery by avoiding exhaustive search over all possible causal structures.
• To provide a practical algorithm that leverages conditional independence tests between disturbances and direct causes to identify the true causal graph.
• To validate the theoretical findings through simulations demonstrating the method's effectiveness under various nonlinear and noise configurations.

Proposed method

• Theoretical analysis of the PNL model using functional equations: X = f(Z) + ε, where f is a nonlinear function and ε is additive noise.
• Identification of non-identifiable cases through enumeration of degenerate functional forms and noise distributions that violate the genericity conditions.
• Derivation of sufficient conditions for identifiability based on the non-Gaussianity and nonlinearity of the error terms and the invertibility of the nonlinear functions.
• Application of the PNL model to each member of the Markov equivalence class of a causal graph to test whether the disturbance is independent of its direct causes.
• Use of conditional independence testing to select the correct causal structure among Markov equivalent graphs, thereby avoiding full enumeration of all possible DAGs.
• Design of a search strategy that prunes the space of candidate structures by leveraging the PNL model's identifiability properties.

Experimental results

Research questions

• RQ1 Under what conditions is the post-nonlinear causal model identifiable in the two-variable setting?
• RQ2 Which specific functional and noise configurations lead to non-identifiability in the PNL model?
• RQ3 Can the PNL model be extended to causal discovery in systems with more than two variables?
• RQ4 Is it possible to avoid exhaustive search over all possible causal graphs by leveraging the PNL model’s structure?
• RQ5 How can the conditional independence between disturbances and direct causes be used to identify the correct causal structure in a Markov equivalence class?

Key findings

• The PNL causal model is identifiable in almost all cases, with non-identifiability occurring only in specific degenerate configurations of the nonlinear functions and noise distributions.
• Sufficient conditions for identifiability are derived, including non-Gaussianity of the noise and invertibility of the nonlinear functions, which ensure that the causal direction can be uniquely recovered.
• In multi-variable settings, the correct causal structure can be identified by testing conditional independence between each variable’s disturbance and its direct causes across the Markov equivalence class.
• The proposed method avoids exhaustive search over all possible causal graphs by leveraging the PNL model’s identifiability, significantly reducing computational complexity.
• Simulations confirm that the method successfully recovers the true causal structure under various nonlinearities and noise types, especially when the noise is non-Gaussian.
• The framework enables reliable causal discovery even in the presence of measurement distortion and inner noise effects, as modeled by the PNL structure.