NB: for the formalism needed to understand what follows please
The large majority of approaches applied so far to estimate TE implicitly follow uniform conditioned embedding schemes where the components to be included in the embedding vectors are selected arbitrarily or separately for each time series. Note that the TE can be seen as a difference of two conditional entropies (CE), or equivalently as a sum of four Shannon entropies:
Taking into account, for instance, the vector approximated using the embedding vector , where and are the embedding dimension and embedding delay, the same for and approximated by and , it is possible to distinguish between a first phase during which the past states are collected and a second phase during which the estimate of the entropy, and consequently of the CE, is evaluated by means of the chosen estimator, according to the following pseudo code:
- build the vectors ;
- use and to evaluate the last two entropies of (1) and, consequently, the lowest CE term (CE2);
- use to evaluate the first two entropies of (1) and, consequently, the highest CE term (CE1);
- compute TE as equal to the difference CE1 – CE2.
The obvious arbitrariness and redundancy introduced by this strategy are likely to cause problems such as overfitting and detection of false influences, Vlachos (2010). Moreover one should assess which TE values are significant. The significance tests associated to TE estimation based on UE are different for model-based and model free estimators.
Nonuniform state-space reconstruction and coupling detection. In: Phys Rev E, 82 (1), pp. 016207, 2010.
Looking for related topics? Search Google