

Overview

Assume we want to compute transfer entropy from a source variable $S$ to a target variable $T$, potentially conditioned on variable(s) $C$. To control the transfer entropy, the analyst must

1. Create a generalised delay reconstruction from the input data. In practice, this means that lagged instances of the time series are organised in some data structure that can be sampled column-wise (e.g. a custom delay reconstruction or a Dataset).
2. Specify a discretization scheme, which can be either rectangular or triangulated.

3. Map variables of the generalised delay reconstruction to the correct marginals. In other words, which columns of the input data should be grouped in which marginals? There are four marginals that must be assigned:

   • The future of the target ($T_{f}$).
   • The present and past of the target variable ($T_{pp}$).
   • The present and past of the source variable ($S_{pp}$).
   • The present/past/future of any conditional variables ($C_{pp}$). If $C_{pp}$ is non-empty, then conditional transfer entropy will be computed. Otherwise, non-conditioned transfer entropy is computed.

The latter is achieved by creating a TEVars instance, which refers to the columns of the input data by their indices.

Once these steps are understood, you're in a position to estimate transfer entropy.