Right. It's a trade-off. OTOH, if you only call one algorithm, you don't want to pay for accumulation that is not used. So maybe we need algorithms that takes iterators and algorithms that take some kind of accumulator object.
Or you could parameterise the accumulator so it only accumuates what you need.
Unfortunately: this method is prone to numerical overflow/underflow :-(
How can this be different than just accumulating it all from scratch? (Or is it the accumulator method in general that is error-prone?)
Yes, the method we're both using I think is the "schoolboy" accumulate the sum of the squares method. It's perfectly good enough for many purposes but Knuth has an alternative that accumulates the sum of the differences from the "working" mean only, see the last method at http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance I haven't investigated this at all except to note that it exists, and actually looking again it might support most of the operations that the current methodology does (not sure about rms mean and the S N-1 "unbiased" variance though). John.