Directional Bias in Discrete Grid Representations

19 concepts21 papers1 proposal

Most computational models represent the world using a grid. Whether it is a lattice of pixels in an image, a volume of voxels in a 3D scan, or a hash table for neural rendering, we rely on axis-aligned structures to organize spatial data. However, these grids are not neutral containers. They possess an inherent "preference" for the horizontal and vertical directions of their own coordinate systems—a phenomenon known as grid-induced anisotropy.

This bias means that an algorithm’s performance often depends on how an object is oriented relative to the grid. As Antonin Chambolle noted in early work on the Mumford-Shah functional, discrete models frequently favor edges that align with the grid axes, leading to "blocky" artifacts and preventing the accurate capture of diagonal or curved geometry.

The Geometry of the Grid

The root of the problem lies in the difference between Euclidean distance and grid distance. In a continuous space, rotating a shape does not change its properties. In a discrete grid, however, the distance between two neighboring points is 1 if they are axis-aligned, but $\sqrt{2}$ if they are diagonal.

This discrepancy ripples through every stage of spatial processing:

  • Segmentation and Boundaries: When algorithms try to find the shortest path or the "perimeter" of a shape, they often "snap" to the grid. Meine et al. and the authors demonstrated that popular watershed algorithms produce jagged, topologically incorrect boundaries because they are forced to follow the underlying lattice.
  • Crystalline Convergence: Mathematically, as grid spacing vanishes, one might expect discrete models to converge to smooth, rotationally invariant results. However, Giacomini and Lussardi proved that discrete energies often converge to "crystalline" perimeters—shapes with flat faces rather than smooth curves—preserving the grid's bias even at infinite resolution.
  • Metrication Artifacts: Viola et al. describe these as "metrication artifacts," where the rectangular structure of pixels biases models toward horizontal and vertical orientations, treating diagonal edges as less accurate approximations.

Anisotropy in Neural Reconstruction

The issue has resurfaced in modern neural graphics. Multi-Resolution Hash Encodings (MHE) are widely used to speed up neural radiance fields (NeRFs), but they rely on axis-aligned grids to store features.

Dai and Fan recently characterized this as a physical system, discovering that MHE is inherently anisotropic. In their analysis, they found that reconstruction is significantly sharper along the primary axes than along diagonal directions. This directional bias leads to visible artifacts in high-frequency regions of a rendered scene, as the grid effectively has a higher "resolution" in some directions than others.

Restoring Rotational Invariance

To combat these artifacts, researchers have developed several strategies to "de-bias" the grid:

  1. Long-Range Interactions: Instead of only looking at immediate neighbors (up, down, left, right), Antonin Chambolle showed that incorporating specific weights for long-range pixel interactions can significantly reduce anisotropy in image segmentation.
  2. Isotropic Regularization: In inverse problems like phase unwrapping, Kamilov et al. proposed using rotation-invariant data-fidelity terms. By penalizing errors globally rather than independently along the H/V axes, they allow for the reconstruction of sharp, arbitrarily oriented edges.
  3. Coordinate Lifting: To handle smooth curves, Chambolle and Pock suggest "lifting" 2D images into a 3D space that includes orientation as a third dimension. This roto-translational framework allows the model to treat all angles equally, enabling the completion of smooth, arbitrary curves.
  4. Structural Tweaks: For neural encodings, Dai and Fan introduced Rotated MHE (R-MHE). By simply rotating the underlying grid structures, they eliminate the alignment between the grid axes and the scene's primary features, effectively smoothing out the directional bias without adding computational overhead.

Ultimately, the goal of this research direction is to move beyond the "pixel-thinking" described by Viola et al. and treat images and volumes as continuous functions, ensuring that the geometry of the math is not dictated by the geometry of the storage.

Key Foundational Work

Finite-differences discretizations of the mumford-shah functional (1999) — contributed 1 concept elements to this direction.
Orientation diffusions (1997) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_grid_induced_anisotropy_0a1c6564,
  title = {grid-induced anisotropy},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/grid-induced-anisotropy/rdr_0a1c656468ae4e49b5e08a6a62cb339b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_grid_induced_anisotropy_0a1c6564_v1,
  title = {grid-induced anisotropy},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/grid-induced-anisotropy/rdr_0a1c656468ae4e49b5e08a6a62cb339b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake