Information transfer in continuous processes

We discuss a recently proposed quantity, called transfer entropy, which uses time series data to measure the amount of information transferred from one process to another. In order to understand its foundation, merits, and limitations, we review some aspects of information theoretic functionals. While for symbol sequences these measures have an intuitive interpretation, their application to continuous state processes and, in particular, their estimation from finite data sets is problematic. For mutual information, finite length scale estimates converge from below and can thus be used to reject the assumption that the observed processes are independent. However, mutual information does not provide any directional information. Conversely, transfer entropy does resolve the directionality of information exchange but no similar monotonic convergence seems to hold. Thus, only in the case of zero transfer entropy in one direction we can reliably infer an asymmetry of the information exchange.