Theoretical Guide

This section provides comprehensive theoretical background for the Causation Entropy library, including mathematical foundations, algorithmic details, and methodological comparisons.

Overview

The Causation Entropy (oCSE) framework provides a principled approach to causal network discovery from time series data using information-theoretic measures. The core philosophy centers on the idea that causal relationships can be quantified through conditional mutual information, which measures the information content shared between variables when conditioning on relevant context.

Mathematical Foundation

The fundamental quantity in causation entropy is the conditional mutual information:

\[I(X_j^{(t-\tau)}; X_i^{(t)} | \mathbf{Z}_i^{(t)}) = H(X_i^{(t)} | \mathbf{Z}_i^{(t)}) - H(X_i^{(t)} | X_j^{(t-\tau)}, \mathbf{Z}_i^{(t)})\]

where: - \(X_j^{(t-\tau)}\) is a potential causal variable at lag \(\tau\) - \(X_i^{(t)}\) is the target variable at time \(t\) - \(\mathbf{Z}_i^{(t)}\) is the conditioning set for variable \(i\)

The “optimal” aspect refers to the systematic selection of the most informative predictors while controlling for statistical significance through permutation testing.

Key Principles

  1. Information-Theoretic Causation: Causal relationships are quantified through information measures that capture statistical dependencies beyond linear correlation.

  2. Forward-Backward Selection: A two-phase algorithm that first selects the most informative predictors (forward) and then removes spurious relationships (backward).

  3. Statistical Significance: All causal relationships are validated through permutation tests to control false positive rates.

  4. Multivariate Conditioning: The framework properly accounts for confounding variables through conditional mutual information.

Next Steps