[time-nuts] Timestamping counter techniques :Estimating accuracy

[Erik Kaashoek]

For a counter without any post-processing the accuracy has a lower bound 
defined by the clock frequency and the gate time. This computed accuracy 
ignores possible error sources outside the counter but at least it is 
easy to calculate and avoids taking too many irrelevant output digits 
into account.
For a counter that uses regression one can compute similar lower error 
bound using also the regression count but it may be possible to 
calculate the RMS of the error over the regression. This seem to require 
storing all the measurement data.
Does anyone know of a method to calculate the RMS error over the whole 
regression in an incremental way without the need to revisit all 
previously stored measurements after the regression has been performed?