Adjacent points in a profile will exhibit a behavior that is analogous to autocorrelation. In case you are not familiar with autocorrelation, it is a time series where the value of a number of subsequent measurements depends on the value of a prior measurement. ... While not true autocorrelation, the analogy should hold true.

This would justify the practice of using a few points to control a surface.

This is a great example of the dire need to develop the

*total variance equation *for each situation. I agree, adjacent points are related. Also, without a doubt, there is autocorrelation from the tool wear (so we can rule out independent, random variation for that participant of variance.) But, as an example, if a position callout is from an inside dimension to an outside datum, the the tool wear on the outside may (and most likely does) wear at a different rate then the tool on the inside. Also, if there is a need to re-chuck to generate one dimension after the other, then the error in chucking will participate (and, it may very well be random and independent).

The worst thing to do is to look at any variation as

*'one output'*. It is a

*multi-modal output*, and how many modes that are significantly influencing the capability will be a function of the CNX analysis - which ones you can adjust - which ones you can keep statistically insignificant, and which ones you have no control over. Those factors also influence the validity of any representation you prepare (capability index, for example) over time. As much as people would love plug and chug - or use it indiscriminately - it is simply not that easy. Some people have proposed using nonlinear response surface modeling to describe the probability of these types of relationships. Sounds interesting, but I have not tried it yet. In any event - one would have to look at each callout, and the processes that generate it, very specifically - it is nearly impossible to generalize.