The 2012 House election numbers sit in the public record like an open question nobody wants to answer. Democratic candidates received 1.4 million more votes than Republicans nationwide. Republicans won 33 more seats. No ballots were destroyed. No precincts were closed early. The votes were counted accurately, every one of them, and the arithmetic produced an outcome that looks like it should be impossible in a functioning democracy.<\/p>\n\n\n\n
The obvious explanation is gerrymandering. That explanation is correct. But “gerrymandering” has become a word that substitutes for understanding \u2014 the name familiar, the mechanism mostly opaque. Because once you see the machinery clearly \u2014 not as a political grievance but as a geometric operation on vote tallies \u2014 you start to understand why the sophisticated tools designed to stop it keep running into the same wall. And why that wall is not a flaw in the tools.<\/p>\n\n\n\n
Start with five hundred thousand voters, split exactly in half: 250,000 for Party A, 250,000 for Party B. The state gets five congressional seats. Five districts of 100,000 voters each. In a system that translates votes into seats proportionally, the natural outcome is somewhere around 3-2.<\/p>\n\n\n\n
Draw Map 1. Three districts go 60,000-40,000 in Party A’s favor; two go 80,000-20,000 against them. Three seats for A, two for B. The map reflects the underlying split.<\/p>\n\n\n\n
Now draw Map 2. Four districts go 55,000-45,000 for Party A. One district goes 90,000-10,000 for Party B. Party A wins four of five seats \u2014 80% \u2014 with exactly half the votes.<\/p>\n\n\n\n
Same voters. Same totals. Different lines.<\/p>\n\n\n\n
The mechanism is visible once you look at where the losing votes went. In Map 2, Party B’s 45,000 supporters in each of the four losing districts contributed nothing to the outcome. Those votes don’t accumulate, don’t transfer, don’t count toward a seat elsewhere. They expire at the district boundary. And in the one district Party B wins overwhelmingly \u2014 90,000 to 10,000 \u2014 more than 39,999 votes were surplus. You win a seat with 50,001 votes; every vote beyond that buys nothing.<\/p>\n\n\n\n
Those are wasted votes. Both kinds: the losing votes cast in a district you never had a chance to win, and the surplus votes piled uselessly onto a win you already secured. A deliberate gerrymander engineers an asymmetry in wasting. Pack your opponents into a few blowout districts where they dominate \u2014 those surplus votes are gone. Distribute their remaining voters thin across many hostile districts where they lose narrowly \u2014 those losing votes are gone too. Meanwhile, spread your own supporters across as many districts as possible, winning each by a comfortable but not extravagant margin. Near zero waste on your side. Enormous waste on theirs. A 50-50 electorate becomes an 80-20 seat advantage. Legally.<\/p>\n\n\n\n
What makes this operational rather than theoretical is that the arithmetic is optimizable. Redistricting software can take precinct-level demographic and voting history data and model, before a single boundary is drawn, which configuration will produce the most asymmetric wasting. Iterate until the map is essentially optimal.<\/p>\n\n\n\n
The arithmetic, worked<\/strong>\n\nMap 2 in the example above produces these totals: Districts 1 through 4 \u2014 Party A 55,000, Party B 45,000 each. District 5 \u2014 Party B 90,000, Party A 10,000. Party B loses four districts and wins one enormous landslide.\nWasted votes, Party B: 45,000 per losing district \u00d7 4 = 180,000 votes spent on losses. In the landslide win, 90,000 \u2212 50,001 = 39,999 surplus votes. Total Party B wasted: approximately 220,000.\nWasted votes, Party A: 10,000 in the one loss, plus roughly 4,999 surplus per winning district \u00d7 4 = 29,996. Total Party A wasted: roughly 30,000.\nParty B wastes seven times as many votes despite identical total vote counts. The efficiency gap \u2014 the difference in wasted votes divided by total votes \u2014 registers at approximately 38% in this fictional state. Real gerrymanders are subtler: the Wisconsin maps that went to the Supreme Court had efficiency gaps of roughly 13% and 10%. The structural logic is identical.<\/code><\/pre>\n\n\n\nThis is not theoretical. In states where Republicans controlled redistricting after the 2010 census, they won approximately 72% of congressional seats while receiving roughly 53% of the votes in 2012, according to the Brennan Center for Justice’s 2013 analysis of post-redistricting congressional control.<\/p>\n\n\n\n
REDMAP \u2014 when the machine was turned on<\/h3>\n\n\n\n
For most of American political history, partisan redistricting was a matter of opportunity. Control the state legislature and governorship in a census year and you drew favorable maps; lose that control and the other party drew them instead. It was cynical and widely understood, bounded by what a roomful of politicians with paper maps and staff could accomplish before the deadline.<\/p>\n\n\n\n
REDMAP ended the era of the paper map.<\/p>\n\n\n\n
The Redistricting Majority Project was the Republican State Leadership Committee’s strategy for the 2010 cycle, built around a simple structural insight: the 2010 midterm elections would determine which party controlled state legislatures when those legislatures drew congressional maps following the census. Win the right state chambers in the right swing states, and you control redistricting in those states for an entire decade. Karl Rove stated this publicly in a Wall Street Journal column on March 3, 2010, under the subheading: “He who controls redistricting can control Congress.” He wasn’t hiding anything.<\/p>\n\n\n\n
REDMAP targeted 107 state legislative races in 16 states \u2014 Wisconsin, Michigan, Ohio, Pennsylvania, and Florida among them. The operation spent approximately $30 million in reported total, though dark money flows through multiple entities make exact totals difficult to verify; the figure appears in ProPublica’s reporting on RSLC and REDMAP funding and in NPR and WBUR’s 2016 coverage, and should be treated as reported rather than precisely audited.<\/p>\n\n\n\n
It worked. Republicans flipped enough state chambers to control redistricting where it mattered. New maps were then produced using statistical and mapping software capable of modeling partisan outcomes at the precinct level \u2014 calculating expected vote-share under a proposed district configuration before any line was finalized. In Wisconsin, the maps that became Act 43 of 2011 were reportedly drafted in a private law firm office rather than a public committee room, specifically to limit disclosure during the drafting process. Court documents describe “new statistical and mapping software”; which tool, exactly, was never publicly disclosed.<\/p>\n\n\n\n
The 2012 results documented the output. Republican House candidates nationally received about 47.6% of the popular vote and won 54% of seats. In states where they controlled redistricting, the conversion was 53% of votes to roughly 72% of seats. The gap between those two ratios is the wasted-vote asymmetry operating at national scale.<\/p>\n\n\n\n
This was not a symmetrical operation on both sides. REDMAP was a nationally coordinated, purpose-built, multi-year strategy targeting 16 states simultaneously, with documented funding, an explicit strategic brief, and post-cycle reporting. There was no Democratic analogue of that scope in 2010.<\/p>\n\n\n\n
What Democrats did<\/strong>\n\nThey gerrymander too, and the examples should be named. Maryland's 6th Congressional District was redrawn in 2011 to transform a reliably Republican seat into a safe Democratic one \u2014 accomplished by attaching suburban Washington D.C. precincts to rural western Maryland and carving away the Republican-leaning communities that had defined the district for decades. Illinois's 4th Congressional District, the \"Earmuffs,\" links two Hispanic communities in Chicago through a thin corridor of highway, producing a majority-minority district that is also reliably Democratic.\nBoth are partisan manipulations and should be called that. Neither constitutes a nationally coordinated, multi-state operation with a documented strategic brief and roughly $30 million in reported funding. REDMAP was qualitatively different in its scope, its explicit planning, and its national execution. Noting the asymmetry is not a partisan claim. It is a description of what the records show.<\/code><\/pre>\n\n\n\nThe court’s long evasion<\/h3>\n\n\n\n
The Supreme Court has been looking at this problem since 1986. The pattern across every case is consistent: the problem is real, it’s harmful, and the Court will not provide a remedy.<\/p>\n\n\n\n
Davis v. Bandemer arrived first, in 1986. Indiana Democrats challenged state legislative maps that gave Republicans a substantial seat advantage despite Democrats winning a statewide vote majority. The Court ruled 6-3 that partisan gerrymandering claims were justiciable under the Equal Protection Clause \u2014 federal courts could hear them. But the majority immediately fractured on what a winning claim would require. Justice White gathered four signatures for an opinion on the merits, not five, and no majority agreed on a standard. The case found Indiana’s maps constitutional and left the threshold for a successful challenge undefined. That is not a resolution; it is a problem deferred.<\/p>\n\n\n\n
Eighteen years later, Vieth v. Jubelirer narrowed the options to two. Pennsylvania Democrats challenged post-2000 congressional maps. Justice Scalia’s plurality of four argued the question was simply nonjusticiable \u2014 a political question courts had no business touching. Four other justices disagreed but couldn’t agree on a standard among themselves. Justice Kennedy, the deciding vote, occupied a position of suspended judgment: the question was justiciable, he believed, but no workable standard had yet been proposed. He left a door open. His concurrence suggested that if a mathematically rigorous, constitutionally grounded standard could be developed, the Court might apply it.<\/p>\n\n\n\n
That concurrence was read, correctly, as an invitation. The efficiency gap was built to answer it.<\/p>\n\n\n\n
Gill v. Whitford in 2018 was supposed to close the loop. Wisconsin plaintiffs challenged Act 43 using the efficiency gap as the constitutional measuring stick \u2014 exactly the kind of standard Kennedy had indicated would matter if one could be found. Nine justices voted unanimously. Not for the plaintiffs. Not against them on the merits. For dismissal on standing. The plaintiffs had argued about statewide partisan injury; the Court held they needed to show district-specific harm to each named plaintiff personally. The efficiency gap as a constitutional test was never evaluated on its substance. Kennedy’s invitation went unanswered.<\/p>\n\n\n\n
He announced his retirement nine days later.<\/p>\n\n\n\n
Rucho v. Common Cause in 2019 closed what Kennedy had left ajar. Chief Justice Roberts, writing for a 5-4 majority, held that partisan gerrymandering claims present “political questions beyond the reach of the federal courts.” Federal courts lacked a “clear, manageable, and politically neutral” standard and therefore could not adjudicate how much partisan advantage crosses a constitutional line. Roberts acknowledged directly that extreme partisan gerrymandering is “unjust” and “incompatible with democratic principles” \u2014 and then ruled it was Congress’s problem, not the Court’s.<\/p>\n\n\n\n
The efficiency gap had been constructed to answer a specific constitutional challenge that a swing-vote justice had placed on the record. By the time the challenge was ready, the justice was gone.<\/p>\n\n\n\n
Where racial and partisan gerrymandering intersect<\/strong>\n\nRacial gerrymandering operates under entirely different law. Thornburg v. Gingles (1986) gave federal courts a functional framework for evaluating racial dilution of minority voting power under the Voting Rights Act: states cannot systematically pack or crack minority voters in ways that reduce their ability to elect preferred representatives. Partisan gerrymandering has no equivalent standard after Rucho.\nThe two categories do not separate cleanly in practice. In states where race and party affiliation are highly correlated \u2014 much of the South, where Black voters vote Democratic at rates above 85% \u2014 drawing \"Republican\" districts and drawing \"white\" districts can be the same geographic act accomplished with the same line. The legal strategy in response is now routine: assert partisan motivation when racial motivation is the challenge, or the reverse, depending on which claim is more legally precarious. The line between permissible partisan manipulation and prohibited racial manipulation is exactly as clear as courts have made it. Which is not very.<\/code><\/pre>\n\n\n\nThe efficiency gap and its rivals<\/h3>\n\n\n\n
Kennedy had asked for a standard. Nicholas Stephanopoulos, a law professor at the University of Chicago, and political scientist Eric McGhee tried to give him one \u2014 a 2015 paper in the University of Chicago Law Review formalising the wasted-vote asymmetry into a single formula.<\/p>\n\n\n\n
The efficiency gap is compact: subtract one party’s total wasted votes from the other’s, divide by total votes cast. Wasted votes are all losing votes in a district, plus all votes beyond the bare majority needed to win for the winning candidate. The result expresses the asymmetry in how efficiently the two parties convert votes into seats.<\/p>\n\n\n\n
Applied to Wisconsin’s Act 43: approximately 13% in the 2012 elections, 10% in 2014. Stephanopoulos and McGhee proposed a threshold: a gap exceeding 7% constitutes presumptively illegal gerrymandering. Act 43 cleared that threshold by a wide margin in both cycles.<\/p>\n\n\n\n
The formula has genuine virtues. It captures the mechanism precisely. It can be calculated from public election results without access to the map-drawers’ intent. It produces a single number a court can evaluate. These were exactly what Kennedy asked for.<\/p>\n\n\n\n
The rivals followed. The mean-median difference measures the gap between a party’s median district vote share and its mean district vote share: when votes are bunched into a few blowout wins, the median drops below the mean. A 2021 paper by Karthik Seetharaman comparing seven gerrymandering metrics across ten elections found the mean-median measure among the least accurate options \u2014 applicable in theory but worse at correctly identifying gerrymandered elections than efficiency gap variants, its values failing to scale with the actual severity of the partisan distortion. Gregory Warrington, a mathematician at the University of Vermont, developed the “declination” metric: the angle between two lines connecting the win rates of districts won by each party. Geometrically elegant. Same underlying problem.<\/p>\n\n\n\n
Every single-number metric is sensitive to where voters happen to live independently of how lines are drawn \u2014 and this is not an edge case. It is the structural condition of American politics in the era of urban concentration.<\/p>\n\n\n\n
Massachusetts provides the clearest demonstration. Democratic voters are so heavily concentrated in Boston and its surrounding suburbs that any congressional map of the state produces lopsided outcomes regardless of intent. A genuinely neutral map \u2014 one drawn with no reference to partisan data, following county lines and optimising for equal population \u2014 might register an efficiency gap that would trigger constitutional scrutiny under the Stephanopoulos-McGhee threshold, not because anyone engineered the result but because of where Massachusetts Democrats happen to live. The state sends nine Democrats and zero Republicans to the House. A 10% efficiency gap produced by deliberate packing-and-cracking is arithmetically indistinguishable from a 10% gap produced by voter geography alone.<\/p>\n\n\n\n
This creates real operational cover. Map-drawers in heavily urbanised states know that natural voter concentration generates a baseline efficiency gap before they’ve moved a single line. Draw a map that scores near that baseline, then argue the gap reflects geography rather than manipulation. The argument is sometimes true. The metric cannot tell the difference.<\/p>\n\n\n\n
The problem that connects every single-number metric runs deeper than any specific formula failure. Each one measures deviation from an implicit baseline \u2014 what a “fair” seat distribution is supposed to look like for the given voter geography. But the baseline isn’t in the data. It’s assumed before the measurement begins: proportionality, partisan symmetry, competitiveness \u2014 each a different political judgment, encoded as a denominator. The choice of metric is the choice of fairness standard, made implicitly rather than stated.<\/p>\n\n\n\n
Wisconsin's numbers<\/strong>\n\nThe Gill v. Whitford lower court record put specific numbers on the Act 43 efficiency gap across two election cycles. In 2012, Republican candidates won 60 of 99 Wisconsin state assembly seats; Democratic candidates won approximately 51.4% of the statewide assembly vote. The efficiency gap registered approximately 13%. In 2014, it was approximately 10%. The plaintiffs argued both figures exceeded the 7% threshold Stephanopoulos and McGhee identified as presumptively unconstitutional, and that the gap's persistence across two election cycles with different turnout patterns indicated it was structural to the map rather than a single-election artifact. The Court's 9-0 ruling on standing meant this argument was never evaluated on its substance.<\/code><\/pre>\n\n\n\nThe ensemble and the arms race<\/h3>\n\n\n\n
Single-number metrics present a practical target beyond their conceptual limits: one number encodes its assumptions visibly enough that an opposing expert witness can spend a trial day dismantling the baseline. The next generation of tools replaced the single number with something harder to attack.<\/p>\n\n\n\n
Moon Duchin, a mathematician then at Tufts University, founded the Metric Geometry and Gerrymandering Group in 2016. The ensemble approach her team developed generates not a score but a distribution. Draw millions or billions of alternative maps satisfying the same legal constraints the enacted map had to satisfy \u2014 contiguous districts, equal population, Voting Rights Act compliance, reasonable compactness \u2014 then ask where the actual map sits in that distribution.<\/p>\n\n\n\n
If 99.9% of randomly generated maps satisfying the same legal requirements give the disadvantaged party more seats than the enacted map does, the enacted map is a statistical outlier. Not unfair by some contested metric \u2014 anomalous among the entire universe of legally valid alternatives. That framing is considerably harder to attribute to natural geography, because the randomly generated maps are drawn from the same state, constrained by the same rules, and they still produce different results from the enacted plan.<\/p>\n\n\n\n
GerryChain, the open-source Python library maintained by MGGG at github.com\/mggg\/GerryChain, implements this using Markov chain Monte Carlo methods \u2014 specifically the ReCom algorithm developed around 2018. Ensemble analysis built on GerryChain was used by expert witnesses in Pennsylvania redistricting litigation in 2018 and has been deployed in multiple state court challenges since.<\/p>\n\n\n\n
What ReCom does<\/strong>\n\nEarlier ensemble approaches generated new maps by flipping individual geographic precincts one at a time between adjacent districts. The problem: the chain moves slowly, staying near the starting configuration rather than exploring the full space of valid maps. For large states, getting a genuinely representative sample was computationally prohibitive. ReCom works differently \u2014 it selects two adjacent districts, merges them into a single region, builds a random spanning tree of that combined region, then cuts the tree at one edge to produce two new valid districts. Each step makes a large structural change, not an incremental adjustment. The chain mixes faster and samples more broadly. This matters because ensemble analysis is only meaningful if the randomly generated maps represent what was actually available to the map-drawer, not just maps that happen to resemble the one already drawn. The method is described in MGGG's 2019 ReCom paper.<\/code><\/pre>\n\n\n\nThe adversarial dimension arrived predictably. The same methodology that identifies outlier maps can construct maps that sit inside the ensemble distribution \u2014 plans that produce preferred partisan outcomes while appearing statistically normal under ensemble testing. Once you know the tool, you can run it yourself before submitting a map, iterating on boundaries until the plan looks like a plausible random sample from the distribution of valid alternatives. Build the detector; use it as a design tool. Whether redistricting vendors have bundled this as a standard commercial feature is less important than the structural reality: the methodology is open and accessible, and the advantage it confers does not depend on any particular vendor. The defenders are not obviously winning.<\/p>\n\n\n\n
And then there’s the constraint problem, which is where the ensemble method runs into the same wall everything else has hit, arrived at more elegantly.<\/p>\n\n\n\n
Generating a random distribution of maps requires specifying which legal constraints to randomise within. Prioritise compactness and you get one distribution; require that county lines be preserved where possible and you get another; interpret VRA compliance strictly and you get a third. Different constraint specifications produce different distributions of random maps, different baselines, and potentially different verdicts on whether the enacted map is an outlier. The result of the test is a function of the setup. Setting up the test requires deciding which values the map should serve \u2014 which is exactly the political question the test was designed to sidestep.<\/p>\n\n\n\n
The baseline problem \u2014 why math can’t answer the question it was asked<\/h3>\n\n\n\n
What does a fair redistricting map look like?<\/p>\n\n\n\n
The question has at least three distinct answers in serious current use, and they are mathematically incompatible with each other.<\/p>\n\n\n\n
Proportionality: a party that wins X% of the votes should win approximately X% of the seats. Competitiveness: as many districts as possible should be genuinely contestable, neither outcome determined before anyone votes. Partisan symmetry: if the parties’ vote shares were exchanged, the seat outcomes would also exchange \u2014 neither side holds a structural advantage that persists regardless of their actual support.<\/p>\n\n\n\n
Choose one and you give up the others. A map optimised for proportionality often requires creating safe seats for each party, which eliminates competitive races. A map optimised for competitiveness can produce non-proportional outcomes when one party’s voters are geographically concentrated \u2014 you get close races, but the seat split diverges from the vote split. Partisan symmetry can be achieved by making every district a safe seat symmetrically, which scores perfectly on the symmetry metric while producing an election in which nothing is actually contested. No realistic map of most states simultaneously satisfies all three.<\/p>\n\n\n\n
This is the crack in everything.<\/p>\n\n\n\n
Every metric embeds one of these definitions as its assumed baseline. The efficiency gap assumes symmetric conversion efficiency \u2014 proportionality-adjacent. The mean-median difference assumes the median district should reflect the mean partisan composition \u2014 also proportionality-adjacent. Symmetry metrics assume symmetry directly. The ensemble method encodes the same judgment in the choice of constraints: which definition of compactness, which VRA interpretation, whether county lines bind. The baseline is never derived from the data. It is chosen before measurement begins, and the choice is the political dispute the measurement was supposed to resolve.<\/p>\n\n\n\n
Chief Justice Roberts wrote in Rucho that federal courts lacked a “clear, manageable, and politically neutral” standard because adjudicating partisan gerrymandering would require courts to “make their own political judgment about how much representation particular political parties deserve.” The criticism of that holding as judicial evasion is fair \u2014 there was a judgment to make and the Court refused. But Roberts was also pointing at something structurally real: the metrics are proxies for contested political values, not measurements of an objective quantity.<\/p>\n\n\n\n
The baseline is a political commitment in arithmetic clothing. Calling the Court cowardly for refusing to pick a proxy doesn’t change what the proxy is.<\/p>\n\n\n\n
Daryl DeFord and Ellen Veomett documented specific mathematical failure modes in symmetry-based metrics in a 2025 Election Law Journal paper, demonstrating that the metrics systematically misidentify partisan gerrymandering in predictable cases depending on the voter distribution. The same year, La Matematica published “Don’t Trust a Single Gerrymandering Metric,” making the structural argument with mathematical proof: single metrics are, as a category, insufficient to the task of detecting gerrymandering because they all require choosing a baseline, and no baseline is derivable from the data. A Scientific American essay titled “Math Can’t Solve Gerrymandering” brought the same conclusion to a wider audience \u2014 evidence that it has escaped the technical literature \u2014 though it is a reflection piece rather than primary evidence for the structural claim, which rests on DeFord\/Veomett and La Matematica.<\/p>\n\n\n\n
The researchers who built these tools have published precise accounts of where they fail and why. That is the most honest thing a researcher can do. It is also a slow-moving acknowledgment of something the baseline problem implied from the start.<\/p>\n\n\n\n
California’s Citizens Redistricting Commission lists its criteria in explicit priority order \u2014 a public encoding of the value judgment rather than an outsourcing of it to a formula. Proportional representation systems remove the leverage entirely by allocating seats by vote share. Both make the political choice explicit rather than embedding it in arithmetic where it becomes invisible.<\/p>\n\n\n\n
The 2012 numbers are unchanged. Democratic candidates, 1.4 million more votes. Republicans, 33 more seats. The mechanism that produced this outcome \u2014 the wasted-vote asymmetry, modelled in software before a single vote was cast, optimised across targeted swing states \u2014 is exactly what the efficiency gap and the ensemble methods were designed to expose. They exposed it well.<\/p>\n\n\n\n
What they could not do is tell you, from first principles, what a fair map looks like. That answer requires a prior commitment to what redistricting is supposed to accomplish, and no algorithm generates prior commitments from data. The metric can identify an outlier. It cannot say whether the outlier is unjust \u2014 because that requires deciding what justice means, which is a question about political values, and always was.<\/p>\n\n\n\n
The mathematical work done on gerrymandering over the past twenty years is not a failure. It is a map of exactly why the problem is hard. The tools showed, with precision, where the political disagreement actually lives. That is not a solution.<\/p>\n\n\n\n
It might be more useful than one.<\/p>\n\n\n\n