Hi everyone,
In case you're interested, with Graham, Joe, and Abhinav, we posted a paper
comparing ReqSketch and t-digest to arXiv:
https://arxiv.org/abs/2102.09299
The upshot is that we can enforce an almost arbitrarily large error in estimates
of t-digest using an adversarial construction, and quite a large error even on
i.i.d. samples drawn from a highly non-uniform distribution. Still, there's no
strong conclusion conclusion as t-digest performs much better on more realistic
distributions, such the uniform or normal distributions. A bit surprisingly, the
DataSketches implementation of ReqSketch is substantially faster than t-digest.
Best regards,
Pavel Vesely
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
