Hacker News Reader

I feel obligated to point out that, by the central limit theorem, it doesn’t matter if the timings are normally distributed. The measurements only need to be independently and identically distributed with finite variance for the estimate of the mean to be normally distributed.

https://en.m.wikipedia.org/wiki/Central_limit_theorem

View on HN

I think he's saying that since the distribution follows a log-normal, the mean is harder to pin down than if it were a normal due to the higher variance. From the CLT, the variance of the mean statistic is var(X)/sqrt(N) -- the var(X) is higher and he'd expect from an assumed normal of X.

The other point he makes is that the mean is not very representative in this wide log-normal scenario, and optimizing for the difference between min and mean is more relevant.

My 2 cents.

Also to nitpick on the article:

"It is not possible, in a normal distribution, to be multiple times the standard deviation away from the mean."

I guess he meant probable :)

One can also work in log space to do the stats, and report the geometric mean in linear space…

I think we're nitpicking too much here, but the main takeaway seems to be to not assume normals and to actually measure things. Seems pretty straightforward advice, but if you've worked anywhere, you'll know how often people content themselves with reporting averages.

Reply on news.ycombinator.com