Ted Dunning

I would add that while I didn't respond well to this issue, but the test `com.tdunning.math.stats.TDigestTest#singleSingleRange` addresses this same issue. ```java public void singleSingleRange() { TDigest digest = factory(100).create(); digest.add(1);...

> This is not so. We consider all ranks. Since the error is largest at the median for the t-digest, this is equivalent to considering only the median. Where an...

> Perhaps we ought to reconsider our method with respect to accuracy measurements. I think that your method (max absolute error) is absolutely fine for evaluating the KLL by itself....

> > benchmarks that show about 50 ns amortized insertion times > > At what input size? > I am not sure if I understand your point here. Do you...

That is much closer to what I saw in my measurements.

It isn't surprising that this changed. I did a fair bit of work adjusting the buffering strategy over the last few years. In particular, the strategy was changed to use...

Hey there. I can't tell exactly what you are doing. Normally what would be done would be to add samples (not centroids) to a t-digest for some short window (1...

Hmm... there are going to be some serious problems here. - when you insert data points with weights, you have a serious risk of violating the internal invariants of the...

I haven't heard back on this in over 6 months and will likely close this issue soon if I don't hear more.

Thanks for your excellent comments. I am re-running the distribution tests to see what has happened with the new algorithms, but I don't have answers on that yet. The quadratic...