A framework for causal discovery in non-intervenable systems.

Area of Science:

• Complex Systems Science
• Information Theory
• Causal Inference

Background:

• Existing causal inference frameworks struggle with complex nonlinear systems.
• A complete theoretical framework for nonlinear causal discovery is lacking.

Purpose of the Study:

• To present a novel, fully nonlinear framework for causal inference.
• To provide a complete information-theoretic disentanglement of causal processes.
• To analyze systems beyond directed acyclic graph representations.

Main Methods:

• Utilizes information-theoretic measures like conditional mutual information.
• Introduces a new concept of 'certainty' to quantify available information.
• Applies the framework to simplified stochastic processes, the Lorentz 1963 system, and El-Nino-Southern-Oscillation data.

Main Results:

• The framework achieves nonlinear causal disentanglement and identifies causal strengths of unknown processes.
• It successfully analyzes systems with nonlinear interactions and those not representable by directed acyclic graphs.
• Demonstrates advantages over existing methodologies in analyzing real-world complex systems like ENSO.

Conclusions:

• The new framework offers a powerful approach to causal discovery in complex nonlinear systems.
• It captures local dynamics and large-scale structures, outperforming traditional methods.
• Information-theoretic measures provide deeper insights than integrated quantities in causal analysis.