Could this just be a result of low sampling by the author? If you reduce a sample to only include some tiny edge case, the resulting data points are going to be weird in random ways.
Edit: Why the downvote? Each of the between-state correlations can be calculated from 40,000 datapoints.