Of course, with a small-enough grid rotational symmetry should be restored (or you've chosen a poor discretization!). One might rephrase your statement as: the lattice artifacts / discretization errors vanish faster on a hex grid. Could be.
When I worked on fire-spreading models [0] the grids we used were square. I would wager that this is because it makes thinking significantly easier. In my current field[1], square lattices are chosen because of the underlying supercomputing architecture. The communication mesh in tera- and petascale computers tend to be rectangular [2]. At these scales, having a nice rectangular layout of the local subvolumes simplifies communication algorithms and can provide a dramatic speedup---in this case having fewer neighbors is better.
[0]: In high school I implemented the fact that fires burn faster uphill while working for David Keyes on http://www.cs.cmu.edu/~caliente/
[1]: Lattice QCD codes use a (4D spacetime) rectangular lattice https://usqcd-software.github.io/
[2]: The BlueGene/Q, for example, is a 5-dimensional rectangular torus.
If you were talking about EE DSP stuff, you'd call it quantization noise. Theoretically a floating point A/D converter would be nice / interesting if such a thing existed, but in practice you minimizing that noise source by using more fixed bits per sample.
In higher dimensions there are applications to broadband coding (you want the densest possible constellation).
Disclosure: I heard some advanced lectures about FEM intended for students who intended to write a diploma thesis about this topic.
http://en.wikipedia.org/wiki/Forest-fire_model
The grid is usually rectangular, and I've always found this a bit odd: hexagons represent 2-d sphere packing, so they always seemed more natural to me. I once asked researchers in the field about this, around 2006, and they responded that rectangular grids serve just as well. I just found this paper from 2007 where apparently hexagonal grids fare better:
http://www.sciencedirect.com/science/article/pii/S0307904X06...
This makes me wonder, for realistic models that are meant to model a notion of neighbour cells, why aren't we always using hexagonal grids in 2d or higher-dimensional analogues? With rectangular grids, you're always faced with the choice of defining whether touching on edges and corners count as neighbours or not, which seems like an unnatural choice. Why, then, does this not seem to matter in the end?