Original post is here. I have changed the order of \(\beta_n\) coefficients.

This also changes the interpretation. What we see here is a pattern that is repeated for each ordering, but the mismatching phase requires significantly fewer states only for the case where there is no break in strategy (i.e. \(\beta_1\) -> \(\beta_4\) -> \(\beta_3 \)).

This is in accordance with the original dataset that suggested breaks in strategy harmed the efficiency of the coding. For the strategy-breaking case, we have a monotonous increase in the number of states required, but only the difference between the first phase and the other two are significant. For the other case, we have an increase when the number of meaning dimensions go up, but a decrease when the number of signal dimensions go down, both significant.

This again underlines that we cannot look at mappings independent of the order of presentation.

The most notable difference between findings here and findings for the original dataset is whether the second matching phase, i.e. 2:2 with order 2, increases or decreases the number of required states, or whether or not \(\beta_4 \gt \beta_1\) to be precise. This tells us that our original dataset is much more compressible during matching phases, and the discrete dataset is more suitable for compression using combinatorial strategies.