One way is to represent squares as a square-that-is-a-power-of-two, plus a “clipping size” that is used when the square is rastorized. If you’re using a lot of optimizations for blank regions, the extra padding has a negligible effect on performance.
I suppose that would work, but all the operations you do on the region quadtree have to be aware of the "clipping size" as well. For example, the rotation transformation that this article implements would also have to rotate that "clipping size" as the unwanted portion of the square is now in a different ___location.
One can use recursive properties better if the structure divides equally. So if not having a strict quadtree is okay, they can divide the square into 3x3, 5x5, etc
Another approach would be to allow a cell's value to be null, meaning it should be clipped, but then you run the risk of having transforms on the region quadtrees that cause the null cells to migrate inwards into the square (such that they cannot simply be clipped off later).