Reproduce “On the Identifiability of the Post-Nonlinear Causal Model” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kun Zhang, Aapo Hyvarinen
The authors analyze when a post-nonlinear causal model can orient two-variable relationships. The model combines nonlinear effects, independent additive disturbance, and invertible measurement distortion; orientation tests whether disturbance is independent of the proposed cause. They characterize five non-identifiable situations and give sufficient identifiability conditions under smoothness and density assumptions. Simulations show both directions can pass; ambiguity prevalence is unknown.
Key insights
- The PNL causal model is generally identifiable, but five specific scenarios, including four novel non-Gaussian cases, render it non-identifiable.
- The derived non-identifiable conditions for the PNL causal model challenge and refine previously established separability theorems for PNL mixing ICA.
- A new theorem demonstrates that in PNL acyclic causal models, mutual independence of disturbances is equivalent to the causal Markov condition combined with local disturbance-parent independence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.