Latent Confounded Causal Discovery via Lie Bracket Geometry

Researchers introduce two novel causal discovery algorithms, BRIDGE and Spectral Kan-Do Flow Matching, that leverage category-theoretic principles and differential geometry to identify causal relationships in systems with latent confounders. The methods reduce the search space for valid causal models by many orders of magnitude while inferring hidden structure directly from intervention-induced geometric flows.

This paper represents a significant theoretical advancement in causal inference by connecting category theory, information geometry, and differential topology to solve a longstanding challenge in causal discovery. The work builds on Kan-Do-Calculus, which treats interventions and conditioning as dual categorical operations, translating this abstract framework into concrete algorithms that can identify causal structures even when confounding variables remain unobserved.

The core innovation lies in recognizing that Radon-Nikodym derivatives between observational and interventional distributions generate vector fields whose failure to close under Lie bracket operations reveals the presence of latent confounding. Rather than treating hidden confounders as intractable obstacles, BRIDGE converts geometric obstructions into actionable signals, using bracket residuals to screen candidate causal edges and propose high-recall families of plausible relationships. This approach dramatically compresses the exponential search space inherent to causal discovery.

For researchers in machine learning and causal inference, this work opens new methodological directions by grounding causal discovery in differential geometry rather than purely statistical criteria. The spectral factorization approach in SKFM enables interpretable learning of intervention dynamics while simultaneously revealing latent curvature structure. The practical implications extend to scientific domains where latent confounding is endemic—epidemiology, economics, and systems biology—where current methods either ignore hidden variables or require strong identifying assumptions.

The experimental validation demonstrates both algorithms successfully recover causal structures with latent confounders while remaining computationally tractable. Future development should focus on handling continuous latent variables, scaling to high-dimensional systems, and clarifying the identifiability guarantees under various distributional assumptions.

• →BRIDGE and SKFM algorithms use Lie bracket geometry to detect latent confounders by analyzing how intervention-induced vector fields fail to close
• →The methods reduce causal discovery search spaces from super-exponential to tractable by leveraging category-theoretic bi-adjunction structure
• →Radon-Nikodym derivatives between observational and interventional measures generate computable geometric obstructions that indicate hidden causal structure
• →Spectral factorization of latent curvature enables amortized learning of intervention fields without explicit confounding variable recovery
• →Approach enables causal discovery in empirical settings with latent confounding, addressing a critical gap in existing methods