Transfer entropy on real, continuous data

Question

Here is the goal and problem: I am trying to calculate a measure of coupling between real-valued, continuous oscillatory data. The data come from two people producing synchronized rhythmic movements. To do this, I have been asked to use Transfer Entropy as a non-parametric measure as such.

The data being worked with is the phase velocities from both people. I have also thought to compare the phase velocity and relative phasing for a TE calculation. My problem (question), which continuous TE estimator should be used: Kernel or KSG? If the Kernel, how should I choose the threshold and Schreiber length? If the KSG method, should I apply the same embedding dimension and delay to both series, or choose them such that each is state-space embedded with its optimal parameters?

Thanks for the help!

Answer 1

I'm pretty sure I answered this query on the mailing list of my JIDT toolkit for transfer entropy, but I'll add something here so there's an answer on this page.

KSG is best of breed these days. You should embed the source and target series individually, optimally, and choose the source-target delay then which maximises the TE.

Regarding phase velocity / relative phase -- you may be interested to see how transfer entropy was computed in our previous work using phase differentials and relative phase in this application to kuramoto oscillators: http://dx.doi.org/10.1109/ALIFE.2011.5954653