Differentiable Causal Discovery of Linear Non-Gaussian Acyclic Models Under Unmeasured Confounding

Yoshimitsu Morinishi, Shohei Shimizu

Abstract: We propose a score-based method that extends the framework of the linear non- Gaussian acyclic model (LiNGAM) to address the problem of causal structure estimation in the presence of unmeasured variables. Building on the method pro- posed by Bhattacharya et al. (2021), we develop a method called ABIC LiNGAM, which assumes that error terms follow a multivariate generalized normal distribu- tion and employs continuous optimization techniques to recover acyclic directed mixed graphs (ADMGs). We demonstrate that the proposed method can esti- mate causal structures, including the possibility of identifying their orientations, rather than only Markov equivalence classes, under the assumption that the data are linear and follow a multivariate generalized normal distribution. Additionally, we provide proofs of the identifiability of the parameters in ADMGs when the er- ror terms follow a multivariate generalized normal distribution. The effectiveness of the proposed method is validated through simulations and experiments using real-world data.

Submission Length: Long submission (more than 12 pages of main content)

Changes Since Last Submission:

We thank the Action Editor for the helpful guidance. We addressed all requested corrections. These edits are strictly formatting and typographic; no scientific content, methods, results, or conclusions were changed.

Equations fitting within text width (Eqs. 8 and 14).
Both displays were reflowed to fit strictly inside the text block by introducing appropriate line breaks and standard display environments. The mathematical content is unchanged.

QED symbol orphan at the top of Page 11.
We adjusted the end of the preceding proof so that the QED symbol appears on the previous page. Page 11 no longer begins with only the QED box.

Table 1 caption citation formatting.
The duplicate parentheses were corrected; the caption now uses the standard single-parenthesis citation. We also audited other captions for the same issue and found no remaining instances.

Half-empty pages (Pages 14 and 16).
We rebalanced the layout by allowing long algorithms to break across pages where appropriate and by adjusting float placement. This removed large white space and produced a consistent page density without altering content.

Code: https://github.com/Yoshimitsu-try/ABIC_LiNGAM

Assigned Action Editor: ~Sergey_Plis1

Submission Number: 4056