Suppose I use Welford's algorithm to compute the standard deviation of multiple sets of values. I only store all the n, mean and M2 results that the algorithm calculates, thus I have these three results per set.
Is it possible to calculate the standard deviation over all samples in all sets, by using only the stored results? Thus I'm looking for a way to combine the n, mean and M2 results. For n and mean this is trivial, but I don't know how to combine the M2 values.
Note: this would be easy when I would use the naive sum of squares algorithm, but wikipedia recommends against using this algorithm in practice.
Note2: I cannot store all values in all sets, I'm writing software that only gets one chance to do something with each value, thus the algorithm needs to be 'incremental' or 'online'.