Schulze method

The Schulze method is a voting system developed in 1997 by Markus Schulze that selects a single winner using votes that express preferences. The method can also be used to create a sorted list of winners. The Schulze method is also known as Schwartz Sequential Dropping (SSD), Cloneproof Schwartz Sequential Dropping (CSSD), the Beatpath Method, Beatpath Winner, Path Voting, and Path Winner.

The Schulze method is a Condorcet method, which means the following: if there is a candidate who is preferred over every other candidate in pairwise comparisons, then this candidate will be the winner when the Schulze method is applied.

The output of the Schulze method (defined below) gives an ordering of candidates. Therefore, if several positions are available, the method can be used for this purpose without modification, by letting the k top-ranked candidates win the k available seats. Furthermore, for proportional representation elections, a single transferable vote variant has been proposed.

The Schulze method is used by several organizations including Debian, Ubuntu, Gentoo, Software in the Public Interest, Free Software Foundation Europe, Pirate Party political parties and many others.

Description of the method

Ballot

The input to the Schulze method is the same as for other ranked single-winner election methods: each voter must furnish an ordered preference list on candidates where ties are allowed (a strict weak order).[1]

One typical way for voters to specify their preferences on a ballot (see right) is as follows. Each ballot lists all the candidates, and each voter ranks this list in order of preference using numbers: the voter places a '1' beside the most preferred candidate(s), a '2' beside the second-most preferred, and so forth. Each voter may optionally:

• give the same preference to more than one candidate. This indicates that this voter is indifferent between these candidates.
• use non-consecutive numbers to express preferences. This has no impact on the result of the elections, since only the order in which the candidates are ranked by the voter matters, and not the absolute numbers of the preferences.
• keep candidates unranked. When a voter doesn't rank all candidates, then this is interpreted as if this voter (i) strictly prefers all ranked to all unranked candidates, and (ii) is indifferent among all unranked candidates.

Computation

Let $d[V,W]$ be the number of voters who prefer candidate $V$ to candidate $W$.

A path from candidate $X$ to candidate $Y$ of strength $p$ is a sequence of candidates $C(1),...,C(n)$ with the following properties:

1. $C(1) = X$ and $C(n) = Y$.
2. For all $i = 1,...,(n-1): d[C(i),C(i+1)] > d[C(i+1),C(i)]$.
3. For all $i = 1,...,(n-1): d[C(i),C(i+1)] \text{≥} p$.

Let $p[A,B]$, the strength of the strongest path from candidate $A$ to candidate $B$, be the maximum value such that there is a path from candidate $A$ to candidate $B$ of that strength. If there is no path from candidate $A$ to candidate $B$ at all, then $p[A,B] = 0$.

Candidate $D$ is better than candidate $E$ if and only if $p[D,E] > p[E,D]$.

Candidate $D$ is a potential winner if and only if $p[D,E] \text{≥} p[E,D]$ for every other candidate $E$.

It can be proven that $p[X,Y] > p[Y,X]$ and $p[Y,Z] > p[Z,Y]$ together imply $p[X,Z] > p[Z,X]$.[1]:§4.1 Therefore, it is guaranteed (1) that the above definition of "better" really defines a transitive relation and (2) that there is always at least one candidate $D$ with $p[D,E] \text{≥} p[E,D]$ for every other candidate $E$.

Example

In the following example 45 voters rank 5 candidates.

• 5 $ACBED$ (meaning, 5 voters have order of preference: $A > C > B > E > D$)
• 5 $ADECB$
• 8 $BEDAC$
• 3 $CABED$
• 7 $CAEBD$
• 2 $CBADE$
• 7 $DCEBA$
• 8 $EBADC$

The pairwise preferences have to be computed first. For example, when comparing $A$ and $B$ pairwise, there are $5+5+3+7=20$ voters who prefer $A$ to $B$, and $8+2+7+8=25$ voters who prefer $B$ to $A$. So $d[A, B] = 20$ and $d[B, A] = 25$. The full set of pairwise preferences is:

Directed graph labeled with pairwise preferences d[*, *]
Matrix of pairwise preferences
d[*,A] d[*,B] d[*,C] d[*,D] d[*,E]
d[A,*] 20 26 30 22
d[B,*] 25 16 33 18
d[C,*] 19 29 17 24
d[D,*] 15 12 28 14
d[E,*] 23 27 21 31

The cells for d[X, Y] have a light green background if d[X, Y] > d[Y, X], otherwise the background is light red. There is no undisputed winner by only looking at the pairwise differences here.

Now the strongest paths have to be identified. To help visualize the strongest paths, the set of pairwise preferences is depicted in the diagram on the right in the form of a directed graph. An arrow from the node representing a candidate X to the one representing a candidate Y is labelled with d[X, Y]. To avoid cluttering the diagram, an arrow has only been drawn from X to Y when d[X, Y] > d[Y, X] (i.e. the table cells with light green background), omitting the one in the opposite direction (the table cells with light red background).

One example of computing the strongest path strength is p[B, D] = 33: the strongest path from B to D is the direct path (B, D) which has strength 33. But when computing p[A, C], the strongest path from A to C is not the direct path (A, C) of strength 26, rather the strongest path is the indirect path (A, D, C) which has strength min(30, 28) = 28.The strength of a path is the strength of its weakest link.

For each pair of candidates X and Y, the following table shows the strongest path from candidate X to candidate Y in red, with the weakest link underlined.

Strongest paths
... to A ... to B ... to C ... to D ... to E
from A ...
A-(30)-D-(28)-C-(29)-B
A-(30)-D-(28)-C
A-(30)-D
A-(30)-D-(28)-C-(24)-E
from A ...
from B ...
B-(25)-A
B-(33)-D-(28)-C
B-(33)-D
B-(33)-D-(28)-C-(24)-E
from B ...
from C ...
C-(29)-B-(25)-A
C-(29)-B
C-(29)-B-(33)-D
C-(24)-E
from C ...
from D ...
D-(28)-C-(29)-B-(25)-A
D-(28)-C-(29)-B
D-(28)-C
D-(28)-C-(24)-E
from D ...
from E ...
E-(31)-D-(28)-C-(29)-B-(25)-A
E-(31)-D-(28)-C-(29)-B
E-(31)-D-(28)-C
E-(31)-D
from E ...
... to A ... to B ... to C ... to D ... to E
Strengths of the strongest paths
p[*,A] p[*,B] p[*,C] p[*,D] p[*,E]
p[A,*] 28 28 30 24
p[B,*] 25 28 33 24
p[C,*] 25 29 29 24
p[D,*] 25 28 28 24
p[E,*] 25 28 28 31

Now the output of the Schulze method can be determined. For example, when comparing A and B, since 28 = p[A,B] > p[B,A] = 25, for the Schulze method candidate A is better than candidate B. Another example is that 31 = p[E,D] > p[D,E] = 24, so candidate E is better than candidate D. Continuing in this way, the result is that the Schulze ranking is E > A > C > B > D, and E wins. In other words, E wins since p[E,X] ≥ p[X,E] for every other candidate X.

Implementation

The only difficult step in implementing the Schulze method is computing the strongest path strengths. However, this is a well-known problem in graph theory sometimes called the widest path problem. One simple way to compute the strengths therefore is a variant of the Floyd–Warshall algorithm. The following pseudocode illustrates the algorithm.

1. # Input: d[i,j], the number of voters who prefer candidate i to candidate j.

2. # Output: p[i,j], the strength of the strongest path from candidate i to candidate j.

3. 

4. for i from 1 to C

5.    for j from 1 to C

6.       if (i ≠ j) then

7.          if (d[i,j] > d[j,i]) then

8.             p[i,j] := d[i,j]

9.          else

10.             p[i,j] := 0

11. 

12. for i from 1 to C

13.    for j from 1 to C

14.       if (i ≠ j) then

15.          for k from 1 to C

16.             if (i ≠ k and j ≠ k) then

17.                p[j,k] := max ( p[j,k], min ( p[j,i], p[i,k] ) )


This algorithm is efficient, and has running time proportional to C3 where C is the number of candidates. (This does not account for the running time of computing the d[*,*] values, which can be implemented in a straightforward way to take time proportional to C2 times the number of voters.)[original research?]

Ties and alternative implementations

When allowing users to have ties in their preferences, the outcome of the Schulze method naturally depends on how these ties are interpreted in defining d[*,*]. Two natural choices are that d[A, B] represents either the number of voters who strictly prefer A to B (A>B), or the margin of (voters with A>B) minus (voters with B>A). But no matter how the ds are defined, the Schulze ranking has no cycles, and assuming the ds are unique it has no ties.[1]

Although ties in the Schulze ranking are unlikely,[2] they are possible. Schulze's original paper[1] proposed breaking ties in accordance with a voter selected at random, and iterating as needed.

An alternative, slower, way to describe the winner of the Schulze method is the following procedure:

1. draw a complete directed graph with all candidates, and all possible edges between candidates
2. iteratively [a] delete all candidates not in the Schwartz set (i.e. any candidate which cannot reach all others) and [b] delete the weakest link
3. the winner is the last non-deleted candidate.

Satisfied and failed criteria

Satisfied criteria

The Schulze method satisfies the following criteria:

Failed criteria

Since the Schulze method satisfies the Condorcet criterion, it automatically fails the following criteria:

Likewise, since the Schulze method is not a dictatorship and agrees with unanimous votes, Arrow's Theorem implies it fails the criterion

The Schulze method also fails

Comparison table

The following table compares the Schulze method with other preferential single-winner election methods:

Schulze Ranked pairs Copeland Kemeny-Young Nanson Baldwin Instant-runoff voting Borda Bucklin Coombs MiniMax Plurality Anti-plurality Contingent voting Sri Lankan contingent voting Monotonic Condorcet Majority Condorcet loser Majority loser Mutual majority Smith ISDA LIIA Clone independence Reversal symmetry Polynomial time Participation, Consistency Resolvability Yes Yes Yes Yes Yes Yes Yes Yes No Yes Yes Yes No Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes No Yes Yes Yes Yes Yes Yes Yes Yes Yes No No Yes Yes No No Yes Yes Yes Yes Yes Yes Yes Yes Yes No Yes No No Yes No Yes Yes Yes Yes Yes Yes No No No Yes Yes No Yes No Yes Yes Yes Yes Yes Yes No No No No Yes No Yes No No Yes Yes Yes Yes No No No Yes No Yes No Yes Yes No No Yes Yes No No No No No Yes Yes Yes Yes Yes No Yes No Yes Yes No No No No No Yes No Yes No No Yes Yes Yes Yes No No No No No Yes No Yes Yes Yes Yes No No No No No No No No Yes No Yes Yes No Yes No No No No No No No No Yes Yes Yes Yes No No No Yes No No No No No No Yes Yes Yes No No Yes Yes Yes No No No No No No Yes No Yes No No Yes No No No No No No No No Yes No Yes No No Yes No No No No No No No No Yes No Yes No Yes Yes No No No No No No No No No No Yes

The main difference between the Schulze method and the ranked pairs method can be seen in this example:

Suppose the MinMax score of a set X of candidates is the strength of the strongest pairwise win of a candidate A ∉ X against a candidate B ∈ X. Then the Schulze method, but not Ranked Pairs, guarantees that the winner is always a candidate of the set with minimum MinMax score.[1]:§4.8 So, in some sense, the Schulze method minimizes the largest majority that has to be reversed when determining the winner.

On the other hand, Ranked Pairs minimizes the largest majority that has to be reversed to determine the order of finish, in the minlexmax sense. [4] In other words, when Ranked Pairs and the Schulze method produce different orders of finish, for the majorities on which the two orders of finish disagree, the Schulze order reverses a larger majority than the Ranked Pairs order.

History

The Schulze method was developed by Markus Schulze in 1997. It was first discussed in public mailing lists in 1997–1998[5] and in 2000.[6] Subsequently, Schulze method users included Software in the Public Interest (2003),[7] Debian (2003),[8] Gentoo (2005),[9] TopCoder (2005),[10] Wikimedia (2008),[11] KDE (2008),[12] the Free Software Foundation Europe (2008),[13] the Pirate Party of Sweden (2009),[14] and the Pirate Party of Germany (2010).[15] In the French Wikipedia, the Schulze method was one of two multi-candidate methods approved by a majority in 2005,[16] and it has been used several times.[17]

In 2011, Schulze published the method in the academic journal Social Choice and Welfare.[1]

Users

sample ballot for Wikimedia's Board of Trustees elections

The Schulze method is not currently used in parliamentary elections. However, it has been used for parliamentary primaries in the Swedish Pirate Party. It is also starting to receive support in other public organizations. Organizations which currently use the Schulze method are:

Notes

1. Markus Schulze, A new monotonic, clone-independent, reversal symmetric, and condorcet-consistent single-winner election method, Social Choice and Welfare, volume 36, number 2, page 267–303, 2011. Preliminary version in Voting Matters, 17:9-19, 2003.
2. ^ Under reasonable probabilistic assumptions when the number of voters is much larger than the number of candidates
3. ^ a b c Douglas R. Woodall, Properties of Preferential Election Rules, Voting Matters, issue 3, pages 8-15, December 1994
4. ^ Tideman, T. Nicolaus, "Independence of clones as a criterion for voting rules," Social Choice and Welfare vol 4 #3 (1987), pp 185-206.
5. ^ See:
6. ^ See:
7. ^ a b Process for adding new board members, January 2003
8. ^ a b See:
9. ^ a b See:
10. ^ a b 2007 TopCoder Collegiate Challenge, September 2007
11. ^ See:
12. ^ a b section 3.4.1 of the Rules of Procedures for Online Voting
13. ^ a b See:
14. ^ a b See:
15. ^ a b 11 of the 16 regional sections and the federal section of the Pirate Party of Germany are using LiquidFeedback for unbinding internal opinion polls. In 2010/2011, the Pirate Parties of Neukölln (link), Mitte (link), Steglitz-Zehlendorf (link), Lichtenberg (link), and Tempelhof-Schöneberg (link) adopted the Schulze method for its primaries. Furthermore, the Pirate Party of Berlin (in 2011) (link) and the Pirate Party of Regensburg (in 2012) (link) adopted this method for their primaries.
16. ^ a b Choix dans les votes
17. ^ fr:Spécial:Pages liées/Méthode Schulze
18. ^ §12(4), §12(15), and §14(3) of the bylaws, April 2013
19. ^ Election of the Annodex Association committee for 2007, February 2007
20. ^ Ajith, Van Atta win ASG election, April 2013
21. ^ §6 and §7 of its bylaws, May 2014
22. ^ §9a of the bylaws, October 2013
23. ^ See:
24. ^ Project Logo, October 2009
25. ^ "Codex Alpe Adria Competitions". 0xaa.org. 2010-01-24. Retrieved 2010-05-08.
26. ^ Civics Meeting Minutes, March 2012
27. ^ Report on HackSoc Elections, December 2008
28. ^ Adam Helman, Family Affair Voting Scheme - Schulze Method
29. ^ See:
30. ^ "Guidance Document". Eudec.org. 2009-11-15. Retrieved 2010-05-08.
31. ^ Democratic election of the server admins, July 2010
32. ^ Campobasso. Comunali, scattano le primarie a 5 Stelle, February 2014
33. ^
34. ^ article 25(5) of the bylaws, October 2013
35. ^ 2° Step Comunarie di Montemurlo, November 2013
36. ^ article 12 of the bylaws, January 2015
37. ^ Ridefinizione della lista di San Cesareo con Metodo Schulze, February 2014
38. ^ article 51 of the statutory rules
39. ^ Voters Guide, September 2011
40. ^ See:
41. ^ GnuPG Logo Vote, November 2006
42. ^ §14 of the bylaws
43. ^ "User Voting Instructions". Gso.cs.binghamton.edu. Retrieved 2010-05-08.
44. ^ Haskell Logo Competition, March 2009
45. ^ "Hillegass-Parker House Bylaws § 5. Elections". Hillegass-Parker House Wiki. Retrieved 25 April 2014.
46. ^ article VI section 10 of the bylaws, November 2012
47. ^ A club by any other name ..., April 2009
48. ^ See:
49. ^
50. ^ Kubuntu Council 2013, May 2013
51. ^ See:
52. ^ article 8.3 of the bylaws
53. ^ The Principles of LiquidFeedback. Berlin: Interaktive Demokratie e. V. 2014. ISBN 978-3-00-044795-2.
54. ^ Lumiera Logo Contest, January 2009
55. ^ bylaws
56. ^ The MKM-IG uses Condorcet with dual dropping. That means: The Schulze ranking and the ranked pairs ranking are calculated and the winner is the top-ranked candidate of that of these two rankings that has the better Kemeny score. See:
57. ^ "Wahlmodus" (in German). Metalab.at. Retrieved 2010-05-08.
58. ^
59. ^ See:
60. ^ bylaws, September 2014
61. ^ 2009 Director Elections
62. ^ NSC Jersey election, NSC Jersey vote, September 2007
63. ^ Online Voting Policy
64. ^ See:
65. ^ National Congress 2011 Results, November 2011
66. ^ §6(10) of the bylaws
67. ^
68. ^ article 7.5 of the bylaws
69. ^ Rules adopted on 18 December 2011
70. ^ Verslag ledenraadpleging 4 januari, January 2015
71. ^ 23 January 2011 meeting minutes
72. ^ Piratenversammlung der Piratenpartei Schweiz, September 2010
73. ^ Article IV Section 3 of the bylaws, April 2013
74. ^ Rules of Procedure, January 2015
75. ^ 2006 Community for Pittsburgh Ultimate Board Election, September 2006
76. ^ §16(4) of the bylaws, November 2014
77. ^ Committee Elections, April 2012
78. ^ LogoVoting, December 2007
79. ^ See:
80. ^ Squeak Oversight Board Election 2010, March 2010
81. ^ See:
82. ^ Election status update, September 2009
83. ^ §10 III of its bylaws, June 2013
84. ^ Minutes of the 2010 Annual Sverok Meeting, November 2010
85. ^ constitution, December 2010
86. ^ article VI section 6 of the bylaws
87. ^ Ubuntu IRC Council Position, May 2012
88. ^ See this mail.
89. ^ Pairwise Voting Results
90. ^ See e.g. here (May 2009), here (August 2009), and here (December 2009).
91. ^ See here and here.
92. ^ Result of Arbitration Committee Elections