While I love seeing mathy articles on HN, the lead image in this article is terrible.
At first glance I see we're talking about numerical integration so I assuming the red part is this method that is being discussed and that it's much better than the other two. Then I look at the axis, which the caption notes is log scale, and see that it goes from 1.995 to 2. Uh-oh, is someone trying to inflate performance by cutting off the origin? Big no-no, but then wait a minute, this is ground truth compared to two approximations. So actually the one on the right is better. But the middle one is still accurate to within 0.25%. And why is it log scale?
Anyway point is, there's lots of room for improvement!
In particular it's probably not worth putting ground truth as a separate bar, just plot the error of the two methods, then you don't need to cut off the original. And ditch the log scale.
At first glance I see we're talking about numerical integration so I assuming the red part is this method that is being discussed and that it's much better than the other two. Then I look at the axis, which the caption notes is log scale, and see that it goes from 1.995 to 2. Uh-oh, is someone trying to inflate performance by cutting off the origin? Big no-no, but then wait a minute, this is ground truth compared to two approximations. So actually the one on the right is better. But the middle one is still accurate to within 0.25%. And why is it log scale?
Anyway point is, there's lots of room for improvement!
In particular it's probably not worth putting ground truth as a separate bar, just plot the error of the two methods, then you don't need to cut off the original. And ditch the log scale.