The Copeland or Llull method is a ranked-choice voting system based on counting each candidate's pairwise wins and losses.
In the system, voters rank candidates from best to worst on their ballot. Candidates then compete in a round-robin tournament, where the ballots are used to determine which candidate would be preferred by a majority of voters in each matchup. The candidate is the one who wins the most matchups (with ties winning half a point).
Copeland's method falls in the class of Condorcet methods, as any candidate who wins every one-on-one election will clearly have the most victories overall.[1] Copeland's method has the advantage of being likely the simplest Condorcet method to explain and of being easy to administer by hand. On the other hand, if there is no Condorcet winner, the procedure frequently results in ties. As a result, it is typically only used for low-stakes elections.
Copeland's method was devised by Ramon Llull in his 1299 treatise Ars Electionis, which was discussed by Nicholas of Cusa in the fifteenth century.[2] However, it is frequently named after Arthur Herbert Copeland, who advocated it independently in a 1951 lecture.[3]
The input is the same as for other ranked voting systems: each voter must furnish an ordered preference list on candidates where ties are allowed (a strict weak order).
This can be done by providing each voter with a list of candidates on which to write a "1" against the most preferred candidate, a "2" against the second preference, and so forth. A voter who leaves some candidates' rankings blank is assumed to be indifferent between them but to prefer all ranked candidates to them.
A results matrix r is constructed as follows:[4] rij is
This may be called the "1//0" method (one number for wins, ties, and losses, respectively).
By convention, rii is 0.
The Copeland score for candidate i is the sum over j of the rij. If there is a candidate with a score of (where n is the number of candidates) then this candidate is the (necessarily unique) Condorcet and Copeland winner. Otherwise the Condorcet method produces no decision and the candidate with greatest score is the Copeland winner (but may not be unique).
An alternative (and equivalent) way to construct the results matrix is by letting rij be 1 if more voters strictly prefer candidate i to candidate j than prefer j to i, 0 if the numbers are equal, and -1 if more voters prefer j to i than prefer i to j. In this case the matrix r is antisymmetric.
The method as initially described above is sometimes called the "1//0" method. Llull himself put forward a 1/1/0 method, so that two candidates with equal support would both get the same credit as if they had beaten the other.[5]
Preference ties become increasingly unlikely as the number of voters increases.
A method related to Copeland's is commonly used in round-robin tournaments. Generally it is assumed that each pair of competitors plays the same number of games against each other. rij is the number of times competitor i won against competitor j plus half the number of draws between them.
It was adopted in precisely this form in international chess in the middle of the nineteenth century.[6] It was adopted in the first season of the English Football League (1888–1889), the organisers having initially considered using a 1/0/0 system. For convenience the numbers were doubled, i.e. the system was written as 2/1/0 rather than as 1//0.
(The Borda count has also been used to judge sporting tournaments. The Borda count is analogous to a tournament in which every completed ballot determines the result of a game between every pair of competitors.)
In many cases decided by Copeland's method the winner is the unique candidate satisfying the Condorcet criterion; in these cases, the arguments for that criterion (which are powerful, but not universally accepted[7]) apply equally to Copeland's method.
When there is no Condorcet winner, Copeland's method seeks to make a decision by a natural extension of the Condorcet method, combining preferences by simple addition. The justification for this lies more in its simplicity than in logical arguments.
The Borda count is another method which combines preferences additively. The salient difference is that a voter's preference for one candidate over another has a weight in the Borda system which increases with the number of candidates ranked between them. The argument from the viewpoint of the Borda count is that the number of intervening candidates gives an indication of the strength of the preference; the counter-argument is that it depends to a worrying degree on which candidates stood in the election.
Partha Dasgupta and Eric Maskin sought to justify Copeland's method in a popular journal, where they compare it with the Borda count and plurality voting.[8] Their argument turns on the merits of the Condorcet criterion, paying particular attention to opinions lying on a spectrum. The use of Copeland's method in the first instance, and then of a tie-break, to decide elections with no Condorcet winner is presented as "perhaps the simplest modification" to the Condorcet method.
Like any voting method, Copeland's may give rise to tied results if two candidates receive equal numbers of votes; but unlike most methods, it may also lead to ties for causes which do not disappear as the electorate becomes larger. This may happen whenever there are Condorcet cycles in the voting preferences, as illustrated by the following example.
Suppose that there are four candidates, Able, Baker, Charlie and Drummond, and five voters, of whom two vote A-B-C-D, two vote B-C-D-A, and one votes D-A-B-C. The results between pairs of candidates are shown in the main part of the following table, with the Copeland score for the first candidate in the additional column.
| A | B | C | D | score | ||
|---|---|---|---|---|---|---|
| A | 3:2 | 3:2 | 2:3 | 2 | ||
| B | 2:3 | 5:0 | 4:1 | 2 | ||
| C | 2:3 | 0:5 | 4:1 | 1 | ||
| D | 3:2 | 1:4 | 1:4 | 1 |
| A | 317:183 | 296:204 | 212:288 | 2 |
|---|
In general, if voters vote according to preferences along a spectrum, the median voter theorem guarantees the absence of Condorcet cycles. Consequently such cycles can only arise either because voters' preferences do not lie along a spectrum or because voters do not vote according to their preferences (eg. for tactical reasons).
Nicolaus Tideman and Florenz Plassman conducted a large study of reported electoral preferences.[10] They found a significant number of cycles in the subelections, but remarked that they could be attributed wholly or largely to the smallness of the numbers of voters. They concluded that it was consistent with their data to suppose that "voting cycles will occur very rarely, if at all, in elections with many voters".
Instant runoff (IRV), minimax and the Borda count are natural tie-breaks. The first two are not frequently advocated for this use but are sometimes discussed in connection with Smith's method where similar considerations apply.
Dasgupta and Maskin proposed the Borda count as a Copeland tie-break: this is known as the Dasgupta-Maskin method.[11] It had previously been used in figure-skating under the name of the 'OBO' (=one-by-one) rule.[5]
The alternatives can be illustrated in the 'Able-Baker' example above, in which Able and Baker are joint Copeland winners. Charlie and Drummond are eliminated, reducing the ballots to 3 A-Bs and 2 B-As. Any tie-break will then elect Able.[12]
Copeland's method has many of the standard desirable properties (see the table below). Most importantly it satisfies the Condorcet criterion, i.e. if a candidate would win against each of their rivals in a one-on-one vote, this candidate is the winner. Copeland's method therefore satisfies the median voter theorem, which states that if views lie along a spectrum, then the winning candidate will be the one preferred by the median voter.
Copeland's method also satisfies the Smith criterion.[13]
The analogy between Copeland's method and sporting tournaments, and the overall simplicity of Copeland's method, has been argued to make it more acceptable to voters than other Condorcet algorithms.[14]
To find the Condorcet winner, every candidate must be matched against every other candidate in a series of imaginary one-on-one contests. In each pairing, each voter will choose the city physically closest to their location. In each pairing the winner is the candidate preferred by a majority of voters. When results for every possible pairing have been found they are as follows:
| Comparison | Result | Winner | |
|---|---|---|---|
| Memphis vs Nashville | 42 v 58 | Nashville | |
| Memphis vs Knoxville | 42 v 58 | Knoxville | |
| Memphis vs Chattanooga | 42 v 58 | Chattanooga | |
| Nashville vs Knoxville | 68 v 32 | Nashville | |
| Nashville vs Chattanooga | 68 v 32 | Nashville | |
| Knoxville vs Chattanooga | 17 v 83 | Chattanooga |
The wins and losses of each candidate sum as follows:
| Candidate | Wins | Losses | Net | r | |
|---|---|---|---|---|---|
| Memphis | 0 | 3 | −3 | 0 0 0 0 | |
| Nashville | 3 | 0 | 3 | 1 0 1 1 | |
| Knoxville | 1 | 2 | −1 | 1 0 0 0 | |
| Chattanooga | 2 | 1 | 1 | 1 0 1 0 |
Nashville, with no defeats, is the Condorcet winner. The Copeland score under the 1/0/-1 method is the number of net wins, maximized by Nashville. Since the voters expressed a preference one way or the other between every pair of candidates, the score under the 1//0 method is just the number of wins, likewise maximized by Nashville. The r matrix for this scoring system is shown in the final column.
In an election with five candidates competing for one seat, the following votes were cast using a ranked voting method (100 votes with four distinct sets):
| 31: A > E > C > D > B | 30: B > A > E | 29: C > D > B | 10: D > A > E |
In this example there are some tied votes: for instance 10% of the voters assigned no position to B or C in their rankings; they are therefore considered to have tied these candidates with each other while ranking them below D, A and E.
The results of the 10 possible pairwise comparisons between the candidates are as follows:
| Comparison | Result | Winner | Comparison | Result | Winner | |
|---|---|---|---|---|---|---|
| A v B | 41 v 59 | B | B v D | 30 v 70 | D | |
| A v C | 71 v 29 | A | B v E | 59 v 41 | B | |
| A v D | 61 v 39 | A | C v D | 60 v 10 | C | |
| A v E | 71 v 0 | A | C v E | 29 v 71 | E | |
| B v C | 30 v 60 | C | D v E | 39 v 61 | E |
The wins and losses of each candidate sum as follows:
| Candidate | Wins | Losses | Net | r |
|---|---|---|---|---|
| A | 3 | 1 | 2 | 0 0 1 1 1 |
| B | 2 | 2 | 0 | 1 0 0 0 1 |
| C | 2 | 2 | 0 | 0 1 0 1 0 |
| D | 1 | 3 | −2 | 0 1 0 0 0 |
| E | 2 | 2 | 0 | 0 0 1 1 0 |
No Condorcet winner (candidate who beats all other candidates in pairwise comparisons) exists. Candidate A is the Copeland winner. Again there is no pair of candidates between whom the voters express no preference.
Since Copeland's method produces a total ordering of candidates by score and is simple to compute, it is often useful for producing a sorted list of candidates in conjunction with another voting method which does not produce a total order. For example, the Schulze and Ranked pairs methods produce a transitive partial ordering of candidates, which generally produces a single winner, but not a unique way of tabulating runner-ups. Applying Copeland's method according to the respective method's partial ordering will yield a total order (topological ordering) guaranteed to be compatible with the method's partial order, and is simpler than a depth-first search when the partial order is given by an adjacency matrix.
More generally, the Copeland score has the useful property that if there is a subset S of candidates such that every candidate in S will beat every candidate not in S, then there exists a threshold θ such that every candidate with a Copeland score above θ is in S while every candidate with a Copeland score below θ is not in S. This makes the Copeland score practical for finding various subsets of candidates that may be of interest, such as the Smith set or the dominant mutual third set.