Research · March 2026 · Active

rhombic

Open-source lattice topology benchmarks. An empirical comparison of the cubic and FCC lattices, published with its costs, null results and corrections.

View project

What it does

rhombic compares two ways of tiling three-dimensional space: the cubic lattice, with 6 neighbours per cell, and the face-centred cubic lattice, with 12 neighbours per cell and rhombic dodecahedra as its Voronoi cells. It measures the difference across independent domains, reports exact numbers and publishes the code that generates them.

The measurements

MetricFCC against cubicScale
Average shortest path30% shorter125 to 8,000 nodes
Graph diameter40% smaller125 to 8,000 nodes
Algebraic connectivity2.3 to 2.5× higher125 to 8,000 nodes
Flood fill reach55% more nodes125 to 8,000 nodes
Nearest-neighbour query speed17% faster125 to 8,000 nodes
Signal reconstruction4 to 10× lower MSE216 to 1,000 samples
Embedding neighbour recall+15 to 26pp at 1 hop125 to 1,000 nodes
Edge costabout 2× more edgesthe price

Under structured, direction-based edge weights the Fiedler ratio rises from 2.3× to 6.1×. That result belongs to Paper 2, which is a working draft.

What the comparison does not show

The table compares two spatially embeddable lattices. It does not make FCC a near-optimal graph in the abstract. The pre-registered equal-edge control shows a degree-preserving random rewire of the same edges beating FCC by 4 to 17× on algebraic connectivity, at the price of wire lengths that grow without bound and rule out local physical wiring. The claim that holds is narrower: among spatially embeddable lattice topologies at matched node count, FCC dominates cubic. Where a task has no spatial metric, the library’s own null results say the geometry does not help.

How it is built

The library generates its own evidence. Every benchmark is reproduced by one command:

pip install rhombic && python -m rhombic.benchmark

Corrections and retractions are tracked in the repository, and analyses of the adapter bank are pre-registered with a completeness interlock in code.

Status of the papers

Paper 1, The Shape of the Cell, is complete and has been neither published nor submitted. Papers 2 and 3 are working drafts. Two later studies, The Gauge Is the Obstacle and Typed State Beats Prose, are released from the same repository.

Where to find it

Python 3.10 or later, MPL-2.0, 827 tests. Built through TASUMER MAF.

python · lattice topology · benchmark · graph theory · signal processing · open source