The Same Ballot Can Produce Three Different Winners. A Bold New Math Wants to Fix That.

Run the numbers the way these mathematicians did and something strange happens: vote three different ways on the same ballot and you can get three different winners. In a single eleven-voter, four-candidate election, candidate A is the Condorcet winner, beating every rival head-to-head. Candidate B is the Borda winner, accumulating the most points. Candidate C wins under instant runoff. Same voters, same preferences, three champions. Democracy, it turns out, does not come with a single answer key — and the messy question of which counting rule deserves to be called fair is quietly becoming one of the most exciting frontiers in mathematics.

That is the argument at the heart of "The New Mathematics of Democracy," a survey by Bailey Flanigan of MIT and Ismar Volić of Wellesley College's Institute for Mathematics and Democracy. The paper is not a celebration of elegant theorems for their own sake. It is a case that the field has outgrown its idealized roots — the abstract world where every voter ranks every candidate completely, where preferences are fixed, where profiles are treated as equally probable. Today's democracy is polarized, weaponized, and administered through software, and the mathematics that once flourished in clean hypotheticals is now being dragged into the real world, where it is producing striking, hopeful results: race-blind maps that protect minorities anyway, voting systems that make gerrymandering nearly pointless, and algorithms that can decide how a city spends its money before a single precinct is drawn.

The Science

Democracy has always been mathematical. Counting votes, apportioning seats, and drawing district lines are arithmetic acts with legal consequences. But for most of its history, the mathematics of democracy lived in a rarefied air. The field known as social choice theory, formalized in the mid-twentieth century, asked deep questions about how individual preferences aggregate into collective decisions — and answered them with theorems of stunning pessimism. The most famous, Kenneth Arrow's Impossibility Theorem of 1951, proved that no voting system can satisfy three basic desiderata at once: that no single voter dictates the outcome (non-dictatorship), that a unanimous preference is respected (Pareto efficiency), and that the ranking of two candidates depends only on how voters rank those two (independence of irrelevant alternatives). Arrow won the Nobel Prize for showing that democracy's wish list is mathematically contradictory.

For seventy years, that theorem set the tone. Social choice theory became a science of limitations, cataloging pathologies, uncovering quirky failures, and exposing openings for strategic manipulation. But it did so in a highly stylized setting: it assumed voters submit complete ballots, that every conceivable preference profile is possible, and that all voters inhabit a single ideological spectrum with single-peaked preferences concentrated around a median.

Flanigan and Volić trace what happened when researchers stopped treating these assumptions as sacrosanct and started treating them as hypotheses to be tested against actual elections. The shift began with synthetic models — spatial models placing voters and candidates in a metric space, impartial culture drawing rankings uniformly at random — which traded "what could happen in principle" for "what is likely to happen." More recent work built on these to create parameterized generative models like Plackett–Luce and Bradley–Terry, which assign latent "strengths" to candidates and generate rankings proportional to them. Crucially, those parameters can now be estimated from real data or chosen with practitioners, letting the models be tuned for turnout, polarization, bloc cohesion, and candidate strength. Software packages like VoteKit and Preferential Voting Tools provide the infrastructure to generate and analyze elections under empirically informed models, turning voting theory from an exercise in logic into a form of computational social science grounded in real elections and large surveys like the American National Election Studies and the Cooperative Election Study.

The empirical turn arrived in force. Researchers assembled a database of roughly 3,000 — and growing — ranked-choice elections from the U.S., Scotland, and Australia, and used it to compare how rules perform in practice, not just in theory. They built synthetic electorates from large surveys, constructing voter distributions from real answers about ideology, party affiliation, and hot-button issues. The result is a voting theory that finally asks the question practitioners care about: which rules actually elect the candidates voters want, in the elections that actually happen?

What They Found

The empirical results are remarkably reassuring — and surprisingly rich. Analysis of the thousands of real ranked-choice elections shows that IRV and Condorcet methods pick the same winner over 99% of the time, and that both almost always select a strong candidate. They dramatically outperform Borda and various alternatives; plurality voting performs the worst of all. The pathologies that animated decades of social choice theory — the spoiler effect, vote splitting, strategic burying — turn out to be vanishingly rare in real elections under IRV, which the authors note "unsurprisingly" is not true of plurality.

But the paper's most consequential finding concerns multi-member districts and proportional representation. The core insight here is that single-winner districts, even in the absence of gerrymandering, systematically dilute votes. The authors cite a stark example: about a third of Massachusetts voters vote Republican, yet they cannot elect a single representative to Congress because they are spread so thinly across the state. Two recent Supreme Court decisions — the 2019 Rucho v. Common Cause ruling that partisan gerrymandering claims are not justiciable, and the 2026 Louisiana v. Callais ruling curtailing Section 2 of the Voting Rights Act — have removed the legal guardrails against partisan map-drawing. That makes structural alternatives newly urgent.

Here the mathematics delivers. Simulations across all 50 states show that multi-member districts using the single transferable vote significantly curtail the possibility of partisan gerrymandering. Even more strikingly, they produce far more proportional representation for racial and ethnic minorities than single-winner districts — and critically, this proportionality holds even without race-conscious line-drawing. Race-blind "neutral" maps perform about as well as maps explicitly optimized for minority representation. An empirical analysis of the 2024 Portland, Oregon STV city council elections illustrated the point in practice: people of color were able to elect candidates of choice in every district, and no single bloc of like-minded voters could sweep the results anywhere.

The catch is institutional, not mathematical. The 1967 Uniform Congressional District Act banned multi-member districts for Congressional elections because bloc plurality voting — where a unified majority can take every seat — allowed a majority to sweep an at-large slate and deny representation to minorities. But the Act, the authors argue, "threw out the baby with the bath water": it discarded the entire multi-member system when the real culprit was the tallying rule. Replace plurality with STV's quota-based counting and the vulnerability the Act feared largely disappears.

Figure 1 crystallizes the second major finding about voter realism. Based on the Cooperative Election Study survey in Michigan, it shows the ideological positions of winners across 100,000 simulated four-candidate elections under Condorcet and IRV. Under a "Theoretical Ideal" model — every voter casts a ballot, fills it out completely, knows exactly where candidates stand — Condorcet methods elect centrist candidates more often than IRV, a result consistent with the median voter intuition. But under a "Most Realistic" model that builds in truncated ballots, abstention, and noise in voters' perceptions of candidates' positions, the difference largely evaporates. The two papers that bookend this finding drew opposite conclusions; the discrepancy, the authors argue, "illustrates the need for importing real-world simulation into research."

Why This Changes Things

The deeper message is that the mathematics of democracy is no longer a spectator sport of impossibility theorems. It has become an engineering discipline — and an unusually consequential one. Consider what the multi-member-district results mean in a post-Rucho, post-Callais America. The federal courts have largely stepped out of the gerrymandering fight. The 2026 mid-decade redistricting frenzy is, by the paper's telling, already underway. If map-drawers can now draw district lines with relatively little judicial oversight, the only durable fix is structural: change the unit of representation so that whatever lines are drawn, they cannot so easily convert a minority's support into zero seats. The finding that race-blind multi-member maps perform as well for minority representation as race-conscious ones is significant precisely because it offers a remedy that may survive legal challenge.

This is also a story about what "fair" means and who gets to define it. Classical theory measured voting rules against abstract axioms. The new work measures representational fidelity — how faithfully a multiwinner system translates the support of political, racial, geographic, and other groups into seats — and has brought metric geometry and other tools to bear on the question. Proportionality in principle is one thing; proportionality in practice, across real districts with real cohesion and real geography, is another. The paper hands the reader a genuinely hopeful empirical claim: the voting systems that match our intuitions about fairness are also the ones that survive contact with reality.

There is a cautionary note buried in the findings, too. The history of ranking rules is littered with advocates promising that a particular method will tame polarization by pulling candidates toward the center — the median voter theorem's promise. The Michigan simulations suggest that promise is fragile: once you model real voter behavior, the centrist advantage of Condorcet over IRV disappears. The lesson is not that Condorcet is bad but that claims about how voting rules shape behavior should be tested against behavioral evidence, not idealized assumptions. "The discrepancy between the conclusions of these two papers," the authors write, "illustrates the need for importing real-world simulation into research."

The participants in participatory budgeting — the paper's second case study — face a different set of questions: how should a finite budget be divided among projects voters select, when projects have costs and voters have preferences, and fairness demands that no group be able to capture everything? The mathematics of "equal shares" and proportionality in resource allocation is the active frontier there, with real cities as the laboratory.

What's Next

The authors close by mapping open frontiers that turn the paper's lessons into a research agenda. Three stand out.

First, voter behavior models remain radically incomplete. Real voters abstain, truncate ballots or fill them in wrong, respond to campaigns, vote strategically, and react to conflicting information; candidates reposition, enter and exit races, and form coalitions. One promising recent direction adapts the "smoothed model" from computer science to ranked voting: profiles are worst-case but perturbed by a small amount of neutrally structured noise, and axiomatic analysis reveals which violations are brittle to disturbance and which are robust — including applications to Arrow's theorem itself. The challenge is to merge this worst-case-with-noise view with data-driven models of actual behavior.

Second, preference formation and opinion dynamics. Classical theory treats voter preferences as fixed inputs to an election, but in reality they emerge through campaigns, media, social networks, and deliberation. Network and dynamical models are beginning to describe polarization, algorithmic recommendation systems, and information flow. Sheaf-theoretic models can represent how people communicate different information in different social contexts — a formalism with an almost eerie resonance for an era of fragmented media ecosystems. The open problem is coupling these evolving-preference models to the voting rules themselves, so researchers can ask how the dynamics of opinion formation translate into democratic outcomes.

Third, voting advice applications. As chatbots increasingly offer on-demand political recommendations, the voter advice pipeline — platforms like StemWijzer in the Netherlands, Wahl-O-Mat in Germany, Smartvote in Switzerland, and Vote Compass across North America — has become a mathematical object in its own right. Recent work studies adaptive questionnaires that choose which question to ask next based on prior answers, trading information elicited against recommendation accuracy. Other work, using Swiss Smartvote data, documents how seemingly small choices — how candidates report positions, what metric and weights the platform uses, which dimensions of disagreement enter the calculation at all — can substantially change recommendations. Since a voting advice application is itself an elicitation and aggregation mechanism, it is a natural subject for social choice-style axiomatic analysis.

None of this work proceeds in isolation. The authors are explicit that mathematical design is only part of the challenge: even the most principled intervention must respond to the institutions it embeds in and the populations it serves. The best work, they argue, is done in partnership with social scientists, legal scholars, practitioners, and public officials. The Institute for Mathematics and Democracy at Wellesley exists precisely to host those collaborations.

The stakes are not abstract. "We now live in an environment shaped by intense polarization, declining trust in institutions, fragmented media ecosystems, politically-driven litigation, compartmentalized information flow, and erosion of longstanding democratic norms," the authors write — while also noting that new systems of participation — ranked choice, proportional representation, citizens' assemblies, participatory budgeting, AI-assisted tools — are moving into practice "often by popular demand from disillusioned citizens." The mathematics described here is not a luxury. It is one of the few tools we have for redesigning the machinery of collective choice so that it can survive the very forces currently testing it. And for the first time in the field's long history, the math is being tested against the democracy it is meant to serve.