Misrepresentation — the measure that hands Borda the trophy, and why that's the lesson¶

A 301 theory page, and the other half of the paper that started distortion. Having proved that no ranked rule is ever perfect, Procaccia & Rosenschein went looking for a model where possibility results exist — and found one by borrowing misrepresentation from Monroe's proportional-representation work: assume each voter's dissatisfaction with a candidate is simply that candidate's rank position on their ballot. Under that assumption the impossibilities lift and every method gets a finite score. Borda scores a perfect 1. That result is worth knowing, and worth being suspicious of, for exactly the same reason — because under this measure the Borda count and the thing being measured are the same arithmetic.

Level: 301 · deep dive Builds on distortion (301) — read that first; this page assumes its vocabulary.

Companions: Same ranks, different utilities — the same paper's impossibility, counted on three ballots · Cardinal utility — whether the numbers assumed here exist at all · When a criterion is built to fit the method — the pattern this page is a case study in · ABC rules & the utilitarian–egalitarian spectrum — where Monroe's idea lives in this repo already.

The move: restrict the utilities until answers exist¶

The distortion framework asks what a ranked rule costs you when voters have arbitrary utilities that merely sum to a fixed total. The answer, in that unrestricted setting, is bleak: no rule is perfect (even at 3 voters and 2 candidates), several familiar rules are unbounded, and computing the loss is NP-complete.

So restrict the utilities. Misrepresentation — Monroe's measure, from his 1995 work on fully proportional representation — says a voter's unhappiness with a candidate is just where that candidate sits on their ballot:

Voter i's misrepresentation of candidate j is μⁱⱼ = (rank of j on i's ballot) − 1. Top choice → 0. Second → 1. Last of m → m − 1. A candidate's total is μⱼ = Σᵢ μⁱⱼ, and the best candidate is the one who minimizes it.

Then, exactly as with distortion, a rule's misrepresentation is the worst-case ratio between the winner's total and the best candidate's total.

Notice what this quietly does. Every voter now has the same utility scale — a permutation of 0, 1, …, m−1 — so the ballot determines the utilities completely, with no slack for the adversary to exploit. The intensity question that drives the whole distortion literature is not answered here; it is assumed away. That is the price of the possibility results below, and it should be quoted with them.

The paper is candid that this is a modelling convenience with a natural home: it works when candidates really are interchangeable-but-for-position — its own example is meeting scheduling, where a schedule conflicting with one of your requirements is genuinely "one worse" than a schedule conflicting with none. Elections between human beings are not obviously that kind of problem.

The results¶

Voting protocol	Misrepresentation	In this library's terms
Borda	1 — optimal	The rank-sum rule; see the catch below
Veto (vote against one)	unbounded	Anti-plurality; not used in practice — the cautionary row
Plurality	= m − 1	Choose-one — loss grows with the field
Plurality with Runoff	= m − 1	Top-two runoff — no better than plain plurality
Copeland — i.e. Ranked Robin	≤ m − 1	Best of the non-Borda rules here
Bucklin	≤ m	The graduated-majority method
Maximin (Minimax)	≤ 1.62 (m − 1)	Worst pairwise defeat, minimized
STV ⇒ RCV-IRV single-winner	≤ 1.5 (m − 1)	Elimination is the costliest ranked count here

Two rows deserve a second look.

Plurality with Runoff buys nothing. Adding a second round to choose-one leaves the bound exactly where it was, at m − 1. A runoff fixes a legitimacy problem (the winner clears 50% of the final pair), not an information problem — the first round still threw away everything below your favorite. Worth remembering whenever top-two runoff is offered as the moderate reform.

Ranked Robin beats instant runoff, again. Copeland's m − 1 sits below STV's 1.5(m − 1), and this is the second framework to order them that way: under metric distortion, Copeland is bounded by a constant (≤ 5) while STV/IRV's bound grows with the field. Two different models, built by different authors a decade apart, on different assumptions — same ranking. When results agree across models that disagree about everything else, that's the closest thing this literature offers to a robust finding. Same ballot, better count.

The catch: the measure is the Borda count¶

Borda's perfect score is not a discovery about Borda. It is arithmetic.

The paper proves it in the sharpest possible form. Proposition 5: a scoring rule has misrepresentation exactly 1 — always elects the misrepresentation-minimizing candidate — if and only if its points are affine in rank, α_l = −a·l + b. That is the definition of Borda (up to shifting and scaling). Rearranged: rules that are Borda score 1; rules that are not Borda score worse than 1; and there is nothing in between.

Look at why. Misrepresentation totals a candidate's rank positions. The Borda count totals a candidate's rank positions, upside down. Minimizing μⱼ and maximizing the Borda score are the same operation on the same numbers. A measure defined as rank-sum will crown the rank-sum rule, and the "proof" of Borda's optimality is a change of sign.

This is the repo's standard debate tell, in a peer-reviewed setting: when someone says "method X uniquely satisfies criterion Y," check whether Y was reverse-engineered from X. Here nobody is cheating — Monroe defined misrepresentation for a different purpose (multiwinner representation), and Procaccia & Rosenschein state the affine characterization openly rather than burying it. But the usable claim is narrow, and "academics proved Borda is the optimal voting method" is not it. The paper's own next sentence pulls the trophy back: Borda is notoriously easy to manipulate.

So the honest one-liner: under a rank-based measure of unhappiness, the rank-based rule is optimal by construction, and the interesting question is what happens to every rule that isn't. The table's other seven rows are the actual content.

Three more findings worth carrying¶

Passing the famous criteria does not bound your loss. The Veto rule satisfies participation, monotonicity, and consistency — three of the most-cited criteria in voting theory — and has unbounded misrepresentation (the paper's Remark 3). A rule can be axiomatically well-behaved and still elect a candidate arbitrarily far from the best one. This is the strongest available argument for the repo's position that criteria tables are a starting point, not a verdict: criteria are pass/fail questions about edge cases, and a bound is a statement about magnitude. They measure different things.

But passing the majority criterion does. Any rule satisfying the majority criterion has misrepresentation ≤ 2(m − 1) (Remark 4). A cheap, broad guarantee, and a small formal echo of the metric argument that majority preference implies something about welfare.

A published erratum, for anyone reading the paper alongside this page. Proposition 6 is printed as "F has unbounded misrepresentation iff α₁ > α₂" — and its own proof establishes the opposite direction. The first half assumes α₁ > α₂ and derives the finite bound (m−1)(2α₁−α₂)/(α₁−α₂); the second half assumes α₁ = α₂ and derives ∞. Corollary 5 settles it: Veto (where α₁ = α₂) is the unbounded one. Read the proposition as bounded iff α₁ > α₂ — a scoring rule keeps a finite bound exactly when it puts some daylight between first place and second. Veto, which pays first and second alike, does not.

Computing it¶

The complexity results run the opposite way from the general model, which is the point of restricting preferences in the first place:

In the general (distortion) model, the core decision problem — MIN-SCORE-MAX-UTIL — is NP-complete, by reduction from Knapsack.
In the misrepresentation model, the analogous problem is solvable in O(n²m²) by dynamic programming, and finding the worst-case losing profile is polynomial too.
For scoring rules with two structural properties the paper names (popular loser and even match — plurality and veto have both), computing the misrepresentation admits an FPTAS: approximable to any ε you like, in polynomial time, via a reduction to cardinality-constrained Knapsack.
Open, and still worth someone's time: characterizing which scoring rules have those two properties, and whether computing distortion for scoring rules is NP-complete in general (the authors conjecture yes; no proof).

What this page does not license you to say¶

"Studies show Borda is the best voting method." No. Under a rank-sum measure of loss, the rank-sum rule is optimal by construction, it is one of the most manipulable methods known, and the paper says so on the same page.
A misrepresentation number for STAR, Score, or Approval. There isn't one. The framework analyzes rules that read rankings; STAR's winner is not a function of the ranked profile at all (what a method reads). Same open gap as on the distortion page — don't paper over it.
"Rankings are fine after all." The possibility results here are bought by assuming every voter's intensities are 0, 1, …, m−1. That assumption is precisely what the distortion model refuses to make, and the three-ballot impossibility shows what it costs when it's false.

Where Monroe's idea already lives in this repo¶

Misrepresentation was invented for multiwinner representation, and that's where it does its most respectable work. Monroe's rule assigns each elected representative a disjoint quota of voters and minimizes the total misrepresentation of the assignment — the natural partner of Chamberlin–Courant, which minimizes it without the quota. Both are in this library already, on the approval side: see the ABC rules spectrum (AV / PAV / CC / Phragmén, verified with abcvoting, which implements monroe and cc directly) and the math behind proportional STAR, where the same NP-hardness drives the same move to greedy sequential rules. Reading this page's single-winner ratios next to that spectrum is the clearest way to see that they are one family: pick an unhappiness function, pick an aggregator, and you have named a rule.

Sources¶

Procaccia & Rosenschein, The Distortion of Cardinal Preferences in Voting (CIA 2006, LNAI 4149, pp. 317–331; Springer) — §4 and its Table 1 are this page; §3 is the distortion page's founding-paper section.
Monroe, Fully Proportional Representation, American Political Science Review 89(4), 1995 — where the misrepresentation measure comes from, and why it is a multiwinner idea first.
Chamberlin & Courant (1983) — the companion objective, and the one this repo teaches on the approval side.
Anshelevich, Filos-Ratsikas, Shah & Voudouris, Distortion in Social Choice Problems: The First 15 Years and Beyond (IJCAI 2021) — the survey; useful here mainly for how little the misrepresentation branch was pursued compared with the metric one.

Lean disclosure: peer-reviewed AI / multiagent-systems research, no stake in the US reform fight — the neutral tier. Its blind spot is stated above and is not small: the model's central assumption (unhappiness = rank position) is the very thing the scores-vs-ranks argument is about, so results derived from it cannot settle that argument in either direction.