The Problem

Miller put heroic efforts into data collection. On multiple occasions, it appears that Miller succeeded in conducting over 70 experimental runs of 20 turns each over a 24 hour period. Miller would have had to have conducted about three runs an hour without sleep and with only short breaks to relieve himself. One has to give him credit for being dedicated...

Nevertheless, despite each single data point representing the average of 20 turns of observation, the data points showed enormous scatter over the course of a day. As seen in Miller's Fig 22, he found it necessary to work with running averages of the data, rather than the data itself.

There are several HIGHLY PECULIAR FEATURES of the running average curves shown in Fig 22 (the same curves also show up later in Fig 26 and 27).

1. The plotted points begin at the beginning of a sidereal day, and end at the end of a sidereal day. 
2. Over the course of a sidereal day, the running averages all begin and end on EXACTLY the same value, implying a perfectly periodic signal. 
3. Surely Miller did not always begin his experimental runs at 0:00 h sidereal time, ending them at 23:59 h sidereal time.
4. When computing a running average, one always has a problem with the endpoints. On the other hand, Miller's running averages completely span the entire 24 h sidereal time periods. 

The Explanation

1. The only way to explain these features of his curves is to presume that Miller looped the data from a day's worth of observations with itself, and computed his running averages across the "seam".
2. There is absolutely no way to know where the "seams" lie in Fig 22. Somewhere in each of his plots must be a pair of points representing the end of one observing session and the beginning of the same session. One point in each of his running averages must represent an improper average of data taken over 24 hours apart.
3. This procedure FORCES the running average curves to display an exact 24 h periodicity, even if none existed originally in the data.

Close examination of Fig 22 reveals that Miller repeats data points at the beginning and end of each curve. Notice below on the left where I have indicated two A's, two B's, two C's, and two D's.

Below on the right, I illustrate for Fig 22d what happens if we eliminate the replicated data points and erase the boldly drawn running average lines, whose computation were based on the false presumption that it was perfectly OK to roll the data points into an infinite loop. One can perform a similar procedure on Fig 22a, b, and c, which all show multiple replicated data points. 

When we erase Miller's seductively drawn running averages and eliminate the replicated points, it becomes suddenly quite difficult to imagine the curves as having 24 hour periodicity.

This is DESPITE the fact that in all likelihood, the replicated end points do not represent the "seam" where Miller falsely joined a data set with itself, but rather represent points that in all probability had maximum correlation with each other in real life. Miller's data was indeed, just that dirty...