Studies of electoral districting report a stubborn fact. When districts are forced to be geometrically compact, meaning round and free of sprawling arms, the party whose voters live spread out in rural areas keeps a structural edge, because the other party's voters are packed into dense cities. Pushing the map toward partisan fairness then requires less compact districts. This trade-off has been measured carefully, but always on one fixed set of building blocks, usually voting precincts. Geographers have long known that statistics computed over areas change when you change how the areas are drawn (the "modifiable areal unit problem"). Nobody has checked whether the fairness–compactness trade-off, and the rural advantage in particular, survives a change of building blocks, or whether it partly reflects the arbitrary scale at which precincts happen to be drawn.
We propose to measure this directly. For each state we build a family of coarser unit systems from the precinct map. Each coarse system is made by merging adjacent precincts, and the systems vary two things independently: how many units there are, and how internally similar each unit is. Every coarse plan is therefore also a valid precinct plan, and vote totals add up exactly at every level. On each system we trace the best achievable fairness at each level of compactness. This separates three explanations for the rural advantage: it comes from the geography of voters itself, from the size of the units, or from whether units mix urban and rural ground. Continuous location embeddings learned from satellite imagery serve as one of several ways to define "similar ground." Topology-constrained clustering, which merges regions only when they touch, is the tool that builds the unit systems.
Core Claim
The fairness–compactness frontier is a property of a voter geography and of its unit system, and the second factor can be measured with a controlled design. Because the unit systems are nested, the precinct-level frontier must weakly dominate every coarser one: coarsening only removes feasible plans. What matters is how much each coarsening costs, and whether that cost depends on unit homogeneity rather than on unit count alone. Our main falsifiable prediction is this. Coarsenings that keep units homogeneous in ground conditions (urban density, land use) will preserve the rural structural advantage almost unchanged down to a few hundred units. Coarsenings of the same size that merge across urban–rural edges will shrink the space of fair plans much faster. If the frontier instead stays flat under all coarsenings, that result is also useful. It would mean the frontiers reported at precinct level are robust to unit choice, and that county-preservation rules, which are themselves a coarsening, cost little fairness.
Where This Sits
The work that motivates the proposal fixes the units and varies the plan. Little House (Seat) on the Prairie uses a simulated-annealing recombination search called GREAT to map the frontier between compactness (Polsby-Popper, the ratio of a district's area to that of a circle with the same perimeter, and convex hull scores) and unsigned partisan bias in Pennsylvania, North Carolina, and Wisconsin. It finds that the rural party keeps a 10.8% advantage in "gerrymandering power" even under strict compactness, and that bias under 2% requires compactness only slightly better than the 2011 enacted plan. All of this is measured on one precinct graph. All Politics is Local defines local fairness and finds that "deviating groups" of cracked voters cluster in cities. It audits on fixed precinct adjacency, and its sensitivity to sparse precinct data is exactly the kind of effect that unit choice can drive. Bringing Spatial Interaction Measures into Multi-Criteria Assessment biases ReCom, a Markov-chain sampler that merges two districts and re-splits them along a random spanning tree, toward edges with high mobile-phone flow on Wisconsin block groups. It reports that compactness and within-district interaction rise together, and it explicitly warns that the result may be specific to Wisconsin's scale. Our proposal does not change the sampler's objective. It changes the graph the sampler runs on, in a controlled way.
Two lines of work supply methods. Connected Subspace Clustering gives IterMerge, a Lloyd-style clustering that repairs connectivity at every iteration. It beats post-hoc repair in 73.75% of 160 configurations, and it shows that unconstrained clustering can shatter into 1,966 fragments for 25 requested clusters. That paper also coarsened its grid for computational reasons without studying what the coarsening cost. We reuse IterMerge to build contiguous coarse units, and we make the cost of coarsening the object of study. SatCLIP gives a roughly one-million-parameter location encoder trained by contrastive learning. It pulls together satellite images and coordinates from the same place, so that locations with similar ground conditions get nearby vectors. Its smoothness knob $L$, the maximum degree of the spherical harmonics used to encode coordinates, trades local detail ($L=40$) against smooth generalization ($L=10$). We use these embeddings as a partisan-blind, globally available description of "what the ground is like" for defining homogeneous units. We do not claim they carry political information.
Some prior work constrains the design. Balanced centroidal power diagrams for redistricting shows that population balance to within one person, with contiguity, is achievable on census blocks in all 43 multi-district states. It also shows that moving between continuous and discrete units needs exact reassignment rather than greedy patching, which tells us not to repair coarse plans greedily. Fair redistricting is hard and the hardness result for connected clustering mean that no frontier we compute is certified optimal on real states. We handle this with nesting checks and small exact instances, described below. Agentmandering notes that districting metrics are often uncorrelated. For that reason we report compactness and fairness on the same underlying precinct geometry at every resolution, so that metric changes cannot pass for frontier changes.
Research Program
The first component is a coarsening generator. Starting from the precinct adjacency graph, which has one node per precinct and an edge between precincts that share a border, it produces contiguous, population-respecting unit systems at target sizes of one-half, one-quarter, one-eighth, and one-sixteenth of the precinct count. The generator can run under several merging rules. Random contiguous merging, done by cutting random spanning trees, is the size-only control. Compactness-driven merging uses a cut-edge criterion. Homogeneity-driven merging runs IterMerge on node features, and the features come from one of three sources: SatCLIP embeddings averaged over each precinct at $L=10$ and at $L=40$, census demographic and density variables, or, in Wisconsin only, the SafeGraph mobility flows used by the interaction-ratio study. We also include one deliberately heterogeneity-maximizing rule, which merges each precinct with its most dissimilar neighbor. Counties form a real-world coarsening wherever county lines are stable. The merging rules never see vote data, so homogeneity-driven coarsening cannot simply encode partisanship.
The second component is a frontier estimator that runs identically on every unit system. It is a tempered ReCom and GREAT-style annealing search with a weighted objective of cut edges and unsigned partisan bias, swept over the weights. It is seeded both randomly and from the projected precinct-level optimum. Every resulting plan is lifted to precincts, so compactness (cut edges on the precinct graph, and Polsby-Popper on precinct geometry) and fairness (unsigned partisan bias, seat share against vote share, local fairness at the 0.5 threshold) are always computed on common ground. For each unit system and each compactness threshold, the summary statistics are the minimum achievable bias and each party's maximum achievable seats, following the gerrymandering-power definition of the Little House paper.
Nesting gives a built-in check on optimizer quality, because the precinct frontier must dominate every coarse frontier. Wherever a measured coarse frontier beats the measured precinct frontier, the precinct search has failed. We then rerun it seeded from the lifted coarse plan and report how often this happens. On the 99-county Iowa graph, where Redistricting Algorithms reports an exact integer-programming minimum of 29 cut edges, and on small synthetic grids, we compare heuristic frontiers against exact integer programs to measure how far the search falls short.
Baselines
The naive reference point is the frontier from a standard, unbiased ReCom ensemble on precincts, together with each state's enacted plan. Any claim about unit effects must show movement larger than the seed-to-seed spread of this ensemble.
The strongest existing method is the precinct-level GREAT frontier of Little House (Seat) on the Prairie. It is the only published trace of the full compactness–fairness frontier on the same states, and it reports a 10.8% rural gerrymandering-power advantage in Pennsylvania and unsigned bias under 2% at compactness near the 2011 plan. Before any coarsening result counts, our estimator must reproduce these Pennsylvania figures on precincts within its own seed variance, using the same election data where it can be obtained. "Beating" this baseline here does not mean a better score. It means showing a statistically reliable change in these two quantities that depends on the unit system, with that change larger than the gap between our estimator and the published numbers.
We also run the power-diagram method of Balanced centroidal power diagrams on the block geometry as a pure-compactness anchor point. This tells us where the maximally compact plan sits on each frontier.
The ablation that isolates our key ingredient holds unit count fixed and swaps only the merging rule: random, homogeneous under each feature source, and heterogeneity-maximizing. If homogeneous and random coarsenings of equal size give the same frontier, then homogeneity is not the mechanism, and the size-only story wins.
Fairness of comparison is enforced in three ways. Every unit system gets the same number of annealing proposals and the same weight sweep. Every system is also run to a fixed convergence diagnostic, since coarse graphs mix faster. Each configuration is repeated over ten seeds.
Why not just…?
Why not just rerun the existing frontier on census blocks instead of precincts? Election returns exist only at precinct level. A block-level run requires splitting votes among blocks by some assumption, such as in proportion to population, and that assumption would then drive the result. Coarsening from precincts keeps votes exact and plans nested. We will still run a block-level experiment with population-proportional vote splitting as a secondary check, and we will report it as conditional on that assumption.
Why not just compare precincts against counties? Counties are included, and they matter for policy. But they mix unit size, internal homogeneity, and history, and their size varies widely across states. A county result alone cannot say why the frontier moved. The random and homogeneity-driven coarsenings at matched county counts are what let us interpret the county result.
Why not just use census demographics instead of satellite embeddings to define homogeneous units? This is a real threat, and we treat it as a competing arm rather than a straw man. If demographic and SatCLIP coarsenings produce the same frontiers, the conclusion is that ground-condition homogeneity matters, whatever its source. SatCLIP's remaining value would then be practical: it needs no census, it uses no political data, and its $L$ parameter gives a controlled smoothness knob. We would state that plainly and not claim a modeling advantage the data do not show.
Why not just argue from nesting that finer units are always at least as good? Nesting guarantees the direction of the effect, not its size. Whether coarsening costs a tenth of a seat or two seats, and whether it removes the rural advantage or preserves it, is the empirical question. It cannot be answered from the inclusion relation alone.
Proposed Experiments
We run the full design on Pennsylvania, North Carolina, and Wisconsin, so that the results can be compared with the Little House paper. We add Michigan and Texas, where All Politics is Local finds that locally fair plans make up under 3% of ensembles, against up to 76.9% in Wisconsin. This contrast lets us test whether local-fairness scarcity is sensitive to unit choice. Wisconsin additionally gets the mobility-based coarsening, and we check whether the reported rise in interaction ratio from 3.77 to 3.90 under flow-biased trees persists on coarse units.
Three results would be decisive. The first is the slope of each party's maximum seats against unit count under random coarsening. A flat slope means the published frontiers are robust. The second is the gap between homogeneous and heterogeneity-maximizing coarsenings at matched size, which is the test of the homogeneity mechanism. The third is the change in local-fairness pass rates across unit systems, which tests whether "deviating groups" at precinct level are partly an artifact of fine, noisy units.
A clean negative result is also worth reporting: frontiers that do not move beyond seed noise under any coarsening down to one-eighth of the precinct count. It would tell reformers that the rural advantage is a property of where voters live, and not of how the map was cut into precincts.
References
- Little House (Seat) on the Prairie: Compactness, Gerrymandering, and Population Distribution
- All Politics is Local: Redistricting via Local Fairness
- Bringing spatial interaction measures into multi-criteria assessment of redistricting plans using interactive web mapping
- Connected Subspace Clustering: Hardness, a Scalable Heuristic, and an Application to Sea Level Geodesy
- SatCLIP: Global, General-Purpose Location Embeddings with Satellite Imagery
- Balanced centroidal power diagrams for redistricting
- Fair redistricting is hard
- Agentmandering: A Game-Theoretic Framework for Fair Redistricting via Large Language Model Agents
- Redistricting Algorithms