Reason behind using Manhattan distance

In this page, I noticed

 using Euclidean distance decreased my score against real-world targets considerably

However, having better score is just a result instead of reason. And I've been thinking about the reason why Manhattan works better for years...

Today, something come to my mind. For faster calculation, most of us use SqrEuclidean instead of real Euclidean. This wouldn't affect the order, but once u use squared distance for gaussian function, boom, the actual distance (to the same degree as the Manhattan one) is squared twice, which actually decreases k size dramatically in some cases.

So could you remember whether your Euclidean version gun is using SqrEuclidean and using that (squared distance comparing to Manhattan) for gaussian, or the correct Euclidean distance is used for gaussian?

That was quite a while ago :-) But I know I tested a lot of different distance functions, including exotic things like multiplicative and log-based, and Manhattan worked best. I'm fairly sure I used Euclidean with a sqrt on the squared distance.

Having a gun that is different from what people expect is helpful, since the tuning they do doesn't affect you as much. This is my guess why Manhattan worked best for me

being different sounds reasonable, since there are plenty of vcs surfers (and vcs is more like euclidean than manhattan imho. btw i’m curious about what log-based distance function is ;)

Log based was something like log(1+abs(a1-b1))

I have 2 hypotheses:

- Manhattan distance is more tolerant to noise than Euclidean distance. Squaring a dimension amplifies noise.

- Curse of dimensionality. Euclidean distance behaves oddly at high dimensions.

Squaring does not affect the order of nearest points, then with knn the same data points should be chosen.

And about noice

euclidean seems to be even better when noice has less energy than the main dimensions.