Applied

Where a sincere voter stops approving

Approval voting asks each voter to tick every candidate they find acceptable, and the most ticks wins. A voter with a ranking votes sincerely by ticking a top segment of it — but the ranking does not say where the segment ends. On the 27 ballots where five classic rules elect five different candidates, some choice of sincere stopping points elects each of the five, the Condorcet loser included. The Condorcet winner can always be elected this way; with three candidates and a large electorate, almost every candidate can; and a voter unsure which race is close can do strictly better by skipping a favourite.

Worth reading first: Five rules and five winners · A lie that pays.

Approval voting is the simplest ballot after plurality. Each voter marks every candidate they approve of, as many or as few as they like, and the candidate with the most marks wins. It is used by several professional societies to elect their officers and in the municipal elections of two American cities, and its advocates make a strong case: it never punishes a voter for marking a favourite who cannot win, it lets a voter support a compromise without abandoning a first choice, and its count is a single tally per candidate that any polling station can add up.

What it does not do is turn a ranking into a ballot. Five rules and five winners took 27 ranked ballots over five candidates and counted them five ways — plurality, Borda, instant runoff, the Condorcet comparison and Coombs’s rule — and found five different winners. An approval ballot cannot be read off a ranking at all, because a voter who ranks D above E above B must still decide whether to approve D alone, or D and E, or the top three. This essay counts what that freedom does on the same 27 voters, and the answer is that it is enough to elect any of the five.

Sincerity does not fix a ballot

Sincere approval ballots can elect any of the five. A: 111113 -> 10,0,7,7,7; B: 323221 -> 8,25,7,12,12; C: 411311 -> 8,12,19,7,12; D: 142434 -> 10,13,12,27,13; E: 231112 -> 8,6,7,7,20
Fig. 1 The 27 ballots read as approval ballots. Each column shows one choice of where each group stops approving, shaded, that makes that column’s candidate the sole winner.

Call an approval ballot sincere for a voter with a ranking if it approves a top segment of the ranking: whenever it approves a candidate, it approves everyone the voter ranks higher. A sincere voter never approves C while withholding approval from B whom they prefer. With five candidates, a sincere voter has four meaningful ballots — approve the top one, two, three or four — plus approving nobody or everybody, which change no one’s chances relative to anyone else’s.

The six groups of the profile are seven voters ranking D E B C A, six ranking A B E D C, five ranking C D B A E, five ranking E B C D A, two ranking A B D C E and two ranking C E A D B. The figure shows, for each of the five candidates, one assignment of a stopping depth to each group that elects that candidate outright. If every group approves only its first choice except the last, which approves three, plurality’s winner A wins with 10 approvals against 7. If the groups stop at depths 3, 2, 3, 2, 2 and 1, B wins with 25. C, the instant-runoff winner, wins with 19 when the D-first group approves four candidates and most others approve one. D wins with 27 — every voter — and E with 20.

The figure shows one set of depths per candidate, with every voter in a group stopping at the same place. The full freedom is larger, since two voters with the same ranking may stop at different depths, and the note’s search runs through all of it: voter by voter, it builds every tally that sincere ballots can produce, and there are 331,161 distinct ones. Every candidate is the unique winner of some of them. So on this profile the ranked ballots do not determine the approval winner in any useful sense. Where voters draw their lines decides the election entirely, and the profile admits a line-drawing for each of the five.

That is a fact about approval voting, not a scandal. The rankings record which candidate a voter prefers to which; an approval ballot records something the rankings do not contain, a cutoff between acceptable and not. Advocates of approval voting say that this extra information is exactly what makes it better — the cutoff measures intensity — and critics say that it makes the outcome depend on a decision that voters will make tactically. The 27-voter profile shows how much is at stake in that decision: everything.

Everyone stopping at the same depth

Every voter approving the same number. t=1: 8,0,7,7,5 -> A; t=2: 8,13,7,12,14 -> E; t=3: 10,25,12,14,20 -> B; t=4: 15,25,21,27,20 -> D
Fig. 2 The same ballots when every voter approves the same number of candidates. One is plurality and elects A; four is voting against the last-ranked and elects D; two elects E and three elects B. C needs voters to stop at different depths.

One family of approval outcomes has a name. If every voter approves exactly their top tt candidates, the rule is tt-approval. With t=1t = 1 it is plurality: each voter approves only a favourite, and A’s eight first places win. With t=4t = 4, out of five, every voter approves everyone but their last choice, so the count is a count of last places subtracted from 27 — the anti-plurality or veto rule — and D wins with 27, because no voter ranks D last. In between, 2-approval elects E with 14 approvals and 3-approval elects B with 25.

So the uniform depths alone reproduce four of the five winners: A as plurality’s, B as Borda’s, D as the Condorcet winner, E as Coombs’s. The fifth, C, is never the winner when everybody stops at the same place, and electing it needs depths that differ by group — the D-first group approving four, most of the others approving one. The figure makes a point that four conditions, and no rule that has all of them made with an impossibility theorem: the family of scoring rules — plurality, tt-approval, Borda, veto — is one family with a dial, and turning the dial moves the winner. Approval voting puts the dial in each voter’s hand separately.

The tt-approval rules also show where approval voting’s results come from. A candidate who is many voters’ first choice and few voters’ second or third does well at small tt and badly at large tt: that is A, with eight first places and twelve last places. A candidate who is rarely anyone’s first choice and rarely anyone’s last does the reverse: that is D, with seven first places and no last places at all. Whether a population of approval voters elects A or D depends on whether they approve narrowly or widely, and nothing in their rankings says which they will do.

Wide wins and narrow ones

The five winners are not equally easy to produce. The search behind the first figure kept, for each candidate, the stopping depths that elect it by the widest margin over the runner-up, and those margins are a rough measure of how much of the electorate has to cooperate. A, plurality’s winner, can at best win by 3 approvals: its eight first places are its whole base, it is ranked last by twelve voters, and the depths that keep everyone else’s tallies low also starve A of second-choice support it does not have. C wins by at most 7. E’s best margin is 12, B’s 13, and D’s 14 — D, approved by all 27 voters when every voter stops at D, can win with the other four candidates held to 13 or fewer.

That ordering is close to the ordering by head-to-head strength. D beats all four rivals, E three, B two, C one and A none; only B and E trade places. It is not a coincidence. A candidate’s best approval tally is the number of voters who do not rank it last, and a rival’s tally can be pushed down to the number of voters who rank the rival above it — so the best margin of a candidate is governed by how many voters prefer it to each rival, which is what the pairwise comparisons count. The fewest swaps to a winner measured a candidate’s distance from being a Condorcet winner by the number of adjacent swaps needed; the approval margin is a cruder measure of the same thing, and on this profile the two agree about who is strongest.

It is also a reminder of what approval voting is for. The rule was proposed as a way to let broadly acceptable candidates win without forcing their supporters to abandon first choices, and the widest wins here belong to the broadly acceptable candidates: D, ranked last by nobody, and B, ranked last by only two voters. Plurality’s winner, the candidate with the most first places and the most last places, is the hardest to elect by approval. If voters’ cutoffs were chosen to express something like acceptability — approve the candidates one would be content with — the profile would tend to elect D or B, and a majority wiser than its members is the argument that an aggregate of many such judgements can be better than any one of them. If cutoffs are chosen any other way, the first figure shows that the profile allows anything.

The Condorcet winner can always win

Approving down to the Condorcet winner. thresholds 142434; tallies 10,13,12,27,13
Fig. 3 Every voter approves the candidates down to D and stops. D is approved by every voter who does not rank it last; each rival only by the voters who prefer it to D, fewer than half.

The majority that goes in a circle introduced the Condorcet winner, the candidate who beats every rival head to head. On this profile it is D, beating A by 17 to 10, B by 14 to 13, C by 15 to 12 and E by 14 to 13. Not every profile has one, but when there is one, sincere approval ballots can always elect it, and the argument is the figure.

Let every voter approve the candidates down to the Condorcet winner and stop — approving it and everyone ranked above it, nobody below. (A voter who ranks it last approves everyone but it, which is still sincere.) Then the Condorcet winner is approved by every voter who does not rank it last. Any rival is approved only by voters who rank the rival above the Condorcet winner, and since the Condorcet winner beats every rival head to head, those voters are fewer than half the electorate. Meanwhile the Condorcet winner is approved by every voter who ranks it above that rival, which is more than half. So it wins.

Here D is approved by all 27 voters and no rival by more than 13. The argument is due to Steven Brams and Peter Fishburn, who made it one of the arguments for approval voting: whatever sincere voters do, the Condorcet winner is among the possible outcomes, and voters who know who it is have a sincere way to elect it. The trouble is the converse. The Condorcet winner is among the possible outcomes, but so, on this profile, is the Condorcet loser — A, who loses to every rival head to head, 10 to 17 against B and 13 to 14 against E — and the first figure elected A as well.

Three candidates, and the share of profiles left open

How often sincere approval leaves everything open. 3: 0.0556 / 0.0000; 5: 0.1620 / 0.1244; 7: 0.3001 / 0.2709; 9: 0.4276 / 0.4023; 11: 0.5340 / 0.5122; 15: 0.6919 / 0.6760; 21: 0.8345 / 0.8249; 25: 0.8906 / 0.8838; 31: 0.9410 / 0.9371; 41: 0.9788 / 0.9772
Fig. 4 Three candidates and every profile of up to 41 voters, each ranking equally likely. The share of profiles in which every candidate can win under sincere approval grows towards one; so does the share in which the Condorcet loser can.

The 27-voter profile was built to make rules disagree, so it might be unrepresentative. With three candidates the question can be settled exactly for every profile, because the possible winners have a clean description. A voter ranking xx above yy above zz has two meaningful sincere ballots: approve xx, or approve xx and yy. To make a candidate cc win, the best that sincere voters can do is this: every voter who ranks cc second approves their top two, every voter who ranks cc first approves only cc, and every voter who ranks cc last approves only their favourite. Then cc is approved by every voter who does not rank it last, and each rival dd by exactly the voters who rank dd first. So cc can win exactly when the number of voters not ranking it last exceeds every rival’s number of first places, and no other set of sincere ballots does better for cc.

That criterion is easy to meet. It compares a count that includes most of the electorate — everyone not ranking cc last, typically two thirds of the voters — with a count of first places, typically a third. The figure applies it to every profile of nn voters, weighting each profile by its probability when every voter’s ranking is drawn at random, the impartial-culture model that how often the majority goes in a circle used to measure cycles. With three voters, all three candidates are possible winners in 5.6% of profiles; with eleven voters, in 53%; with 41 voters, in 97.9%. The share in which the Condorcet loser can win, among profiles that have one, rises the same way, to 97.7%. And in no profile at all is the Condorcet winner excluded, as the previous section proved.

So the 27-voter profile is not special. For large electorates and three candidates, sincere approval voting with a random profile is compatible with every outcome nearly always, and the winner is decided by where the voters stop. The same freedom underlies the arguments over district maps in every map of twenty-five voters, where one fixed set of voters gave anything from one seat to four depending on where lines were drawn. Models in which voters set their cutoffs by some rule — approve every candidate above the average of their utilities, say — produce a determinate winner, but the determinacy then comes from the rule and not from the rankings.

The voter who should skip a favourite

When the best approval ballot skips a favourite. ac:0.650 a:0.575 acd:0.575 c:0.550 abc:0.550 ad:0.500 none:0.475 ab:0.475 cd:0.475 abcd:0.475 bc:0.450 d:0.400 abd:0.400 b:0.375 bcd:0.375 bd:0.300
Fig. 5 A voter ranking a > b > c > d who believes either that c and d are tied for the lead or that a and b are. The single best of the sixteen possible ballots approves a and c and skips b.

All of this assumed sincere ballots. Approval voting’s best-known strategic property is that, with complete information, sincerity costs nothing: if a voter knows exactly how everyone else has voted, then among the best ballots for that voter there is always a sincere one. The figure’s note checks this over every combination of the other voters’ totals up to four approvals each, with four candidates and ties broken by lot, and finds no exception. That is part of why approval voting is so often described as hard to manipulate, compared with the profitable lies a lie that pays found under ranked rules.

The guarantee fails when the voter is uncertain. Take a voter who ranks four candidates a>b>c>da > b > c > d, valuing them at 1, 0.6, 0.3 and 0, and who thinks two situations are equally likely: either cc and dd are tied for the lead with aa and bb far behind, or aa and bb are tied for the lead with cc and dd far behind. In the first situation the voter’s ballot matters only through whether it approves cc and not dd; in the second, only through whether it approves aa and not bb. The ballot that does the right thing in both is {a,c}\{a, c\} — approving the favourite and the third choice, skipping the second. Its expected value is 0.650, and every other ballot does worse; the best sincere ones, {a}\{a\} and {a,b,c}\{a, b, c\}, reach 0.575 and 0.550.

So a voter facing two possible close races should vote insincerely, and the reason is structural rather than a quirk of the numbers. An approval ballot is a list of yes-or-no answers, one per candidate, and each answer matters only in the race where that candidate is pivotal. When two different races might be the close one, the voter wants to answer each correctly, and the correct answers need not form a top segment of the ranking. With three candidates this cannot happen, since the only pair a sincere ballot can split are first and second or second and third, and the corresponding races do not conflict; Brams and Fishburn showed that with three candidates every ballot that is not dominated by another is sincere. With four or more, the figure shows that an insincere ballot can be the only best one.

What the counts cannot show

Every figure here is about which outcomes are possible, and none is about which outcome happens. Real approval voters do not choose their stopping points to make a particular candidate win; they follow habits, polls and party signals, and studies of elections that used approval ballots have generally found most voters approving one or two candidates. A count of possible winners says that the ranked ballots alone cannot predict the result; it does not say the result is arbitrary, any more than the possibility of every outcome makes a coin toss unfair.

The impartial-culture weighting also deserves a warning. Treating every ranking as equally likely and independent across voters is the standard way to count profiles, and how often the rules find the majority favourite used it for the same reason, but real electorates are strongly structured: voters cluster around parties and positions, and many of the profiles that dominate the count here — those with three candidates in a near three-way tie — are rare in practice. Under structured preferences the set of possible winners is usually smaller. The figure measures the freedom that approval ballots add to rankings, under the assumption that the rankings carry as little structure as possible.

Finally, approval voting’s merits do not rest on rankings at all. Its defenders argue that voters’ preferences are not rankings but evaluations — this candidate acceptable, that one not — and that an approval ballot records them directly. On that view, the indeterminacy measured here is an artefact of starting from rankings, and the honest question is whether voters’ cutoffs are meaningful. That question is about people, and the figures cannot answer it.

Still open: what voters will do with the dial

Approval voting is in use, and its effects are still being measured. Its two city-wide adoptions in the United States, and its long use in professional societies, give some data on how many candidates voters approve and how the results compare with the plurality outcome; the studies so far report approval and plurality agreeing in most contests. What is not settled is the equilibrium question: if voters in repeated elections adapt their cutoffs to polls, as a chain of majorities leads anywhere found that majorities can be led around by agenda, does approval voting converge to the Condorcet winner when one exists? Jean-François Laslier showed that in one model of large electorates, with voters reacting to uncertain polls, it does, and other models give other answers. The theorem that sincere ballots can always elect the Condorcet winner shows that it is reachable; whether real voters reach it is an empirical question about how people use a dial they have been handed.

Shares its objects with

Essays that name at least two of the same things, and that neither author linked.

Named objects

A dashed tag is an object no other essay names yet.

Borda countCondorcet winnerDominant strategyExhaustive searchStrategyStrategy-proofnessVoting rule