Normalization factor in Uzal-costfunction might be wrong.

[hkraemer]

Hey all,

take a look here. This seems to be valid. The correction also follows from Eqs. (17), (18) & (20) in the original publication. Thanks to @JayeshMD , who digged deeper into the math. However, this will change the general offset of the L-function and Fig.7 in the original paper would look qualitatevily the same (even better), but the graphs have a different offset. The change in the code is subtle, but it does make sense from a physical perspective I'd say. I also tried around with L-functions of different time series lengths of the same system and got similar results as @JayeshMD has shown in his slides: With high time series lengths the value of L saturates/converges. This does not happen the way it is indicated in the paper and also implemented here so far.

Another remark regarding PECUZAL: PECUZAL is not affected by these changes, since there we look at ratios of L for adjacent embedding dimensions and, thus, the time series length cancels out.

I am already preparing a PR and will adjust the tests, which I compare with the outcomes of the PECUZAL-implementations in MATLAB and Python. But I can not use the Figures in the original publication as a benchmark anymore. So let me know, if you are fine with that. While doing that I will also adjust the tests for PECUZAL, regarding #95 .

Cheers