commented:
I remember my mind being blown by this central idea:

this is why this library is so nice, the Hilbert curve is a
space-filling curve that converts multiple dimensions into one
dimension that has an special spatial feature: it preserves the
locality.

I came across this idea in a design doc for Google Maps or Earth back
in 2005 or 2006.  I wasn’t working on those products, but the design
docs were openly shared (I think there was even a central mailing list
for them), and I had a lot of fun reading them.
What I found amazing was that (afaik) the fractal scheme of the
Hilbert curve is the only way to turn 1D cache locality into 2D cache
locality, independent of scale.
Computer memory of course has 1D cache locality.  But on a map, you
want to zoom in, and pan, at multiple scales:

your house
your city
your state
your country

etc.
So if you think about it, if you have terabytes of image tiles of the
entire earth, ordering the 2D tiles along a 1D Hilbert curve is the
most logical way to arrange them.
You don’t want to spray requests for hundreds of tiles across random
servers.  Especially when you consider that they can’t be stored in a
single computer.

Later I think I found some instances of the Hilbert curve in computer
graphics.  Maybe it is not that novel an idea …
Wikipedia mentions this fact -
https://en.wikipedia.org/wiki/Hilbert_curve#Applications_and_mapping_algorithms
And it cites a 2001 paper -
https://ieeexplore.ieee.org/document/908985
So I can imagine in 2005-2006, this is still quite novel
At least back then, I remember my mind being blown … I have never
worked in graphics or maps, so maybe other people’s minds are not
blown.  But I’d be interested in any related work, what has happened
since 2005, etc.

  commented:
It is indeed a very powerful idea!

What I found amazing was that (afaik) the fractal scheme of the
Hilbert curve is the only way to turn 1D cache locality into 2D cache
locality, independent of scale.

There are other curves that preserve some degree of locality, such as
Z-order curves: more discontinuous than Hilbert, but simpler to
compute. Very slightly better than Hilbert might be Moore curves,
which are basically Hilbert curves arranged so the endpoints connect.
(I don’t know what curve is optimal at preserving locality, but I
wouldn’t be surprised if something from the Hilbert family was it.)
.