Path 1 — Physical Length L (= cut score on this grid)
Path 2 — Density-Weighted Length L̃
Population Density (fBm) — brighter = denser
Both path maps start from the same density field and the same random
Voronoi seeds, and advance in the same iteration loop. The only
difference is the objective:
Path 1 (left) uses uniform line tension — population density
enters only through the equal-population constraint — so it minimizes
physical boundary length L (the paper's primary objective).
Path 2 (middle) uses density-weighted diffusion, so it minimizes a
density-weighted boundary length L̃, penalizing boundaries more heavily in
populated areas.
Cut score. On this uniform 4-adjacency grid every shared boundary
segment has the same physical length, so the reported L is simply the count
of grid edges separating different districts — which is exactly the
Duchin–Tenner cut score. Physical length and cut score therefore coincide here
up to a constant scale factor (the common edge length): this is the discrete face of
Remark 4.3, and it is why a separate "cut" map would reproduce Path 1 exactly. Note
that in the 4-adjacency setting L is a Manhattan (staircase) perimeter —
a diagonal interface is charged its horizontal-plus-vertical extent, over-measuring
true Euclidean length by up to √2. The three weightings separate only once edges
carry unequal physical lengths (an irregular census-block mesh, or a richer
neighbourhood); Path 2's density weighting is the one genuine departure on this grid.
Each district is shown in its own color (no political lean). The density panel is a
heat map — brighter means more people per cell. Click and drag on
either map to manually gerrymander that map.