Computational redistricting exercise
246 districts, zero state lines
The United States redrawn as 246 equal-population districts, built algorithmically from 2023 Census county data with no regard for the 50 state borders — and no human hand on any boundary.
The map
Each shape is one district of roughly equal population. Colors are assigned only so that no two touching districts share a color — they don’t encode any category. Hover a district for its stats.
This version optimizes for equal population — an ±8.9% deviation from target, tighter than the geometric version’s ±16.3% — by letting 121 of 246 districts cross a state line. The cost is shape: a median compactness of 0.26, well below the geometric version’s 0.54, because a boundary chasing equal population county-by-county has to bend wherever the population does.
This version optimizes for simple, compact shapes and respecting state lines — zero of 246 districts cross a border, and median compactness is 0.54, roughly double the no-state-lines version’s 0.26 — by accepting looser population equality. Small states (Wyoming, Vermont, New Mexico, Nebraska, Idaho, West Virginia) each get one whole district regardless of whether that leaves it over or under the national target, because splitting them across a state line was the one thing this version wouldn’t do. Neither version is “more correct”: equal population, compact shape, and intact state lines are three separate goals that don’t converge onto a single ideal map, and real redistricting fights are ultimately arguments about which of those to prioritize.
Bottom-left insets: Alaska and Hawaii, each held to a single district — every mainland county they could join sits across open ocean, so splitting either would mean gerrymandering around water, not around people.
Every district here is drawn from whole counties within a single state, cut by straight lines chosen to balance population while keeping the resulting shape as simple as possible — the “shortest splitline” approach. The cost: small states (Wyoming, Vermont, New Mexico…) get one whole district each, even where that’s noticeably over or under the national target, because splitting them across a state line was the one thing this version won’t do.
How it was drawn
The goal was to remove every subjective choice a human map-drawer normally makes. Population and geography are the only inputs — no party registration, no incumbency, no demographic targeting.
-
01
Population & geometry, county by county
2023 Census Bureau county population estimates joined to current county boundaries (3,144 counties, including Connecticut’s 2022 planning regions and Alaska’s current boroughs) — the finest unit that keeps the problem computationally tractable nationwide.
-
02
Split the 16 counties bigger than a district
Los Angeles, Cook, Harris, Maricopa and 12 others each outweigh the 1,361,443-person target on their own. Each was cut into equal-area pieces (assuming uniform population density within the county) so no single district is forced to be one county times a multiplier.
-
03
Recursive population bisection
The remaining counties are repeatedly split in two along whichever geographic axis divides their combined population most evenly, each half kept to a single contiguous landmass via the Census Bureau’s county-adjacency graph, recursing until 244 contiguous pieces remain.
-
04
Balance polish
A boundary-swapping pass nudges counties between neighboring districts wherever it strictly reduces total population deviation, run to convergence — guaranteed to terminate because it never makes the total imbalance worse.
-
05
Alaska & Hawaii, standalone
Neither can be merged into a mainland district without crossing open ocean, so each is simply given one district of its own — Hawaii’s five island counties linked only across their interisland channels.
What “least gerrymandered” means here
No political, racial, or incumbency data ever entered the process — only population counts and which counties physically touch. That’s the strongest, most falsifiable definition of “not gerrymandered” available: a human never chose a line.
- County-level, not block-level. The finest unit is a county (or a piece of one), not a Census block. A block-level run would tighten population balance further but takes far more compute.
- Uniform-density assumption. The 16 split counties are cut by area, not by their true internal population distribution, since sub-county population data wasn’t part of this run.
- Alaska sits at 0.77× target. Unavoidable — it’s an island by land-adjacency terms, so it gets exactly one district regardless of population.
- Sparse-county districts read as less “compact.” The lowest Polsby–Popper scores belong to districts stitching together 50–120 low-population Western counties to reach target — a real consequence of uneven population density, not a sign of manipulation.
- Cartographic boundaries. Shapes use the Census Bureau’s 1:20,000,000 generalized boundary file, appropriate for a national overview map, not survey-grade precision.