Z Curve

The Z curve is a line which zig-zags through a space, going through all points in a particular order. It’s a nice, quick way to enumerate points in a way which sort-of preserves locality, or ‘closeness’, between points in the space and points on the resulting line. Its locality isn’t as good as the Hilbert curve’s, but the Z curve is embarassingly simple to understand:

Given a point with coordinates (x, y), we can find out its position on the Z curve by writing x and y in binary, interleaving their bits, then reading back the resulting number.

For example, the point ( 5 , 10 ) can be written in binary as ( 0101 , 1010 ). If we interleave these, we get 0 1 1 0 0 1 1 0, which converted to decimal is 102.

To convert back to (x, y) coordinates we just read the even and odd bits, respectively.

The interleave function will interleave lists of Bits whilst the outerleave function will extract lists of Bits from a single list:

interleave :: [Bits] -> Bits
interleave l = case l of
                    []        -> []  -- No lists to interleave
                    []    :_  -> []  -- We abort as soon as an empty list is spotted
                    (x:xs):ys -> x : interleave (ys ++ [xs]) -- Extract x, send xs to the end

outerleave :: Int -> Bits -> [Bits]
outerleave n xs = let (heads, xs') = splitAt    n xs          -- Take one bit for each list
                      tails        = outerleave n xs'         -- Recurse to get the rest
                      lists        = zipWith (:) heads tails  -- Prepend the heads to their tails
                   in if length xs < n
                         then replicate n []                  -- Base case
                         else lists

For example, we can colour our Z curve with a black-to-white gradient, then zig-zag it across a 2D space:

f x y = adjust . fromBits $ interleave [toBits y, toBits x]

Since the interleaved number has twice as many bits, we must adjust it to fit in the 128 range we’re using:

adjust = (`div` (2 ^ scale))

This results in the following image:

/build/panpipe-6d6443872c0c404d/bw

Notice that the gradient is quite smooth: large discontinuities are rare, common discontinuities are small. That’s because the curve has good locality.

Higher Dimensions

Since the Z curve works for any number of dimensions, we can zig-zag it through the 3D RGB colour space, where the (x, y, z) coordinates we interleave correspond to red, green and blue colour intensities respectively.

Like before, we interleave the x and y coordinates of our pixels to get their position on the Z curve:

f x y = rgbCube $ interleave [toBits y, toBits x]

To get our red, green and blue values, we just ‘outerleave’ three numbers from this position:

rgbCube xs = let [r, g, b] = outerleave 3 xs
              in (adjust r, adjust g, adjust b)

Again, we adjust the colours to fit in our range. This time it’s not strictly necessary; since they’re only using 2/3 of the allowed bits, all values will be in range, but they’ll be quite dark so we scale them up:

adjust = let cubeScale = floor $ fromIntegral (scale * 2) / 3
             ratio     = 2 ^ (scale - cubeScale)
          in (* ratio) . fromBits

This gives us the following 2D embedding of the RGB colour cube:

/build/panpipe-6d6443872c0c404d/rgb

This isn’t as smooth as the greyscale, since the problem is harder: we don’t have the luxury of an extra dimension, like the greyscale example. Instead, we have to remove a dimension from the RGB cube. Many discontinuities follow the same pattern as the greyscale example, eg. the image being split into quadrants. The tartan-like banding is due to the cube’s neighbours having to rearrange themselves in the constrained 2D space.