causal-automata

causal-automata is a JAX-accelerated cellular automaton engine for causal discovery: recovering a Directed Acyclic Graph (DAG) from observational data by evolving an adjacency matrix under physics-inspired local rules.

Alongside constraint-based search (PC) and continuous score-based methods (GOLEM), the engine treats edges as particles on a grid. Gravity, V-wells (collider sinks), birth, and death update every candidate edge in parallel, with wall times reported under Benchmarks.

See Installation and Quickstart to get started.

User guide

• Installation
  • Requirements
  • From PyPI
  • From a source checkout (contributors)
  • JAX accelerators
  • Build the docs locally
• Quickstart
  • Minimal discovery loop
  • State variables
  • NumPy vs JAX
  • Synthetic end-to-end harness
• Command-line tools
  • Synthetic DAG harness
  • Benchmark vs PC / GOLEM
• Benchmarks
  • Reproduce
  • Interpreting metrics

Theory

• Theory: causal discovery as automata
  • Correlation vs. causation
  • Independence, forks, and colliders
  • From Pearl to grids
  • Generative topology and Bayesian automata
  • Parallel physics as an alternative to sequential search
• Physics rules and matrix mechanics
  • Paper-level forces
  • Implemented tick (code)
  • GPU matrix computation
• Glossary

Reference

• API reference
  • Package exports
  • NumPy Bayesian CA
  • JAX core
  • Synthetic experiments
  • Benchmarks
• Bibliography
  • Related background
  • Software

Indices and tables

• Index

• Module Index

• Search Page