The median voter theorem rests on specific assumptions. When they break down — more parties, multi-dimensional policy, primaries, credibility — so does the prediction of convergence. Explore each limit interactively.
Models and their limits. Every formal model makes simplifying assumptions. The Downsian model's assumptions are what generate its clean prediction of convergence. When we relax them — one at a time — the model yields richer, messier, but more realistic predictions. Click each assumption to see what changes.
Click each assumption to see what happens when it fails
A1
Two parties only → Three or more parties
›
With two parties, the only path to victory runs through the median. With three or more parties, a centrist move risks being outflanked: a party that moves to position 0 can be undercut by a rival who stays at the extreme, capturing the base while the centrist gets outgunned.
⟶ Convergence fails. Parties may stay differentiated. A small party can win by carving out an extreme niche — as Greens, far-right, or regional parties often do.
A2
One dimension → Multi-dimensional policy
›
Real policy covers economics, social issues, foreign policy, environment, identity. In two or more dimensions, a single "median" in all dimensions simultaneously is extremely unlikely to exist. The McKelvey–Schofield Chaos Theorem shows that without a multi-dimensional median, majority voting can cycle through virtually any policy outcome.
⟶ No stable equilibrium. An agenda-setter who controls the order of votes can potentially steer outcomes almost anywhere in the policy space.
A3
Full flexibility → Primaries and credibility
›
Parties must first win a primary election among their own (more extreme) voters before facing the general electorate. The median of the left primary is to the left of the general median; the median of the right primary is to the right. Candidates must satisfy both electorates — a two-stage problem. Even if a party reaches the centre, voters may not believe it will govern from there.
⟶ Platforms stay more differentiated than the pure Downsian model predicts. The "pivot to the centre" after primaries may be partial or disbelieved.
A4
Office-seeking → Policy-seeking parties
›
If parties genuinely care about what policies are implemented — not just whether they win — they face a trade-off: moving to the median increases the chance of winning, but if they do win, they implement a policy far from their preferred outcome. Policy-motivated parties accept some electoral risk to stay closer to their ideal point.
⟶ Parties differentiate. Left parties stay left of the median; right parties stay right. Convergence is partial, not complete. This matches what we observe in practice.
Real voters do not vote purely on ideological proximity. They respond to valence (perceived competence, leadership quality, integrity), partisan identity, retrospective evaluations ("how has the economy done?"), and strategic voting. A voter may prefer a small party but vote for a larger one to avoid "wasting" their vote.
⟶ Party positions are only one of many electoral inputs. A competent centrist and an extreme populist may get the same vote share despite very different platforms.
A6+7
Perfect info + fixed preferences → Imperfect info & persuasion
›
If voters have imperfect information, they may not know where parties truly stand — opening space for strategic ambiguity. If campaigns shift preferences, parties have incentives to move voters toward them rather than moving to the current median. Both relaxations give parties more strategic latitude.
⟶ Parties may benefit from vagueness (appealing to multiple voter types simultaneously) or from investing in persuasion rather than convergence.
What the Downsian model still gives us. Even when its assumptions are violated, the model provides a baseline: it tells us what rational parties would do under ideal conditions. Deviations from convergence — which we observe everywhere — then need to be explained by referencing the specific assumptions that are failing.
No incentive to moderate. The electorate below is polarised: most voters cluster at the left and right extremes, with few in the centre. Party A anchors the left base, Party C anchors the right base, and Party B sits in the middle. See what happens to vote totals when B moves toward the centre — and why no party has an incentive to moderate.
Voter distribution (polarised — peaks at left and right)
Party positions — drag to reposition
Plurality rule: most votes wins. Each voter supports their nearest party.
--
Party A
--
Party B
--
Party C
--
Winner
Loading...
Guided scenarios — click to load
Why no party has an incentive to moderate. With a polarised electorate, the centre is a voter desert. Any party that moves toward 0 loses its dense base to the party that stays extreme, while gaining almost nothing from the sparse middle. The equilibrium is for each party to stay at its voter concentration — not at the median.
Adding a second dimension. Now voters and parties have positions on two policy axes: economic policy (left–right) and social policy (authoritarian–libertarian). Drag voters and parties in 2D space. See why finding a stable "median" in two dimensions is so much harder than in one.
Two-dimensional policy space — drag voters and parties
Drag any voter (circles) or party (squares) to reposition. Each voter votes for the nearest party. Parties are coloured by who controls them.
—
Party A votes
—
Party B votes
—
Leader
Key insight — no unique median. In one dimension, there's always a unique median. In two dimensions, the median in one dimension may not be the median in the other. A "two-dimensional median" (Condorcet winner in 2D) exists only under very special conditions — when voter ideal points have perfect radial symmetry around a centre point (Plott's Theorem, 1967). In practice, this condition almost never holds, and majority cycling can range across the entire policy space.
What stabilises real politics? Despite the chaos theorem, real politics is not random. Stability comes from institutions (procedural rules, constitutional constraints), party discipline (reducing the effective dimensionality of conflict), status quo bias (changing policy is costly), and electorates where ideal points happen to be relatively clustered (a "small yolk").