Multidisciplinary Workshop on Advances in Preference Handling: Papers from the AAAI-14 Workshop Justified Representation in Approval-Based Committee Voting Haris Aziz Toby Walsh NICTA and UNSW, Sydney 2033, Australia NICTA and UNSW, Sydney 2033, Australia Abstract ity’ (Laslier and Sanver 2010). However for the case of multiple winners, the merits of AV are ‘less clear’ (Laslier and Sanver 2010). In particular, for the multi-winner case, AV does not address concerns such as proportional representation. For example, if 51% of the agents approve the same k candidates, and the rest of the 49% agents another disjoint set of k candidates, then the agents in minority do not get any of their approved candidate selected! We consider approval-based committee voting in which agents approve a subset of candidates and based on the approvals, a set of winners is selected. We propose a natural axiom called justified representation that is a strengthening of unanimity. The axiom ensures that a group of agents that is large enough should have at least one approved candidate in the winning set. The concept has some better computational, fairness, and proportionality properties than some of the previous approaches for proportional voting using approval ballots. We show that some standard approval-based multi-winner rules — AV (approval voting), SAV (satisfaction approval voting) and MAV (minimax approval voting) — do not satisfy justified representation. We suggest natural modifications of MAV as well as PAV (proportional approval voting) that satisfy justified representation. We then present a simple linear-time algorithm that satisfies justified representation and constitutes a reasonable approval-based rule in its own right. Finally, we show how justified representation can be used to formulate other attractive approvalbased multi-winner rules. Over the years, multi-winner rules based on approval ballots have been introduced in the literature (Kilgour 2010). In the case of Proportional Approval Voting (PAV ), one reduces the weight of an approval from a particular agent based on how many other candidates the agent approves of have been elected. Another way to modulate the approvals is through computing a satisfaction score for each agent based on their ratio of approved candidates appearing in the committee to their total number of approved candidates (Satisfaction Approval Voting (SAV )). Finally, an approval-based rule that received considerable attention in the literature is MAV (minimax approval voting). Minimax approval voting selects the set of k candidates that minimizes the maximum Hamming distance from the submitted ballots. All the rules informally described above have a more egalitarian objective than AV . Steven Brams the main proponent of AV in single winner elections has argued for example that SAV is more suitable for more equitable representation in multiple winner elections (Brams and Kilgour 2010). Based on their relative merits, approval-based multi-winner rules have been examined in great detail in both economics and computer science in recent years (Brams and Fishburn 2007; LeGrand, Markakis, and Mehta 2007; Meir, Procaccia, and Rosenschein 2008). The Handbook of Approval Voting discusses various approval-based multi-winner rules including SAV , PAV and MAV (Kilgour 2010). However, there has been limited axiomatic analysis of the rules from the perspective of representation. Approaches to generalize approval voting to the case of multiple winners each have their own benefits and drawbacks. One way to compare the various rules in a principled manner is by an axiomatic analysis. Introduction The aggregation of preferences is a central problem in artificial intelligence (Conitzer 2010). In this paper we consider the setting in which agents express preferences over candidates and the goal is to select a specified number of candidates that are deemed winning. Such multi-winner settings cover domains such as committee selection, parliamentary elections, the hiring of faculty members or automated agents deciding on a set of plans (Elkind, Lang, and Saffidine 2011; LeGrand, Markakis, and Mehta 2007). In particular, we focus on approval-based multi-winner rules in which each agent approves of (votes for) a subset of the candidates instead of submit a full-blown ranking over the candidates. Eliciting approvals is simple and has less information requirements than eliciting rankings. The most straightforward way to aggregate approvals is to have every approval for a candidate contribute one point to that candidate and select those candidates that get the highest number of approvals. This rule is called Approval Voting (AV ). AV has many desirable properties in the single winner case (Brams, Kilgour, and Sanver 2006), including its ‘simplicity, propensity to elect Condorcet winners (when they exist), its robustness to manipulation and its monotonic- In this paper, we formalize justified representation in approval-based voting. We show that some prominent rules including AV , SAV , and MAV do not satisfy justified representation. We show that MAV and PAV can be suitably modified so as to satisfy justified representation. We 20 as |Q \ T | + |T \ Q| = |Q| + |T | − 2|Q ∩ T |. Minimax approval voting was proposed by Brams, Kilgour, and Sanver (2007). It is known that computing the outcome of MAV is NP-hard (LeGrand, Markakis, and Mehta 2007). then present linear-time algorithms to verify and compute outcomes that satisfy justified representation. Finally, we show how justified representation can be used to formulate other attractive approval-based multi-winner rules. The area of multi-winner approval voting is related to the study of proportional representation when selecting a committee (Procaccia, Rosenschein, and Zohar 2008). However, the results in the proportional representation literature have largely been negative even for approval ballots (Procaccia, Rosenschein, and Zohar 2008). In this paper, we show that justified representation seems to be a meaningful property for approval-based rules: although it gives justifiably fair solutions, unlike previous approaches, it is not so constraining as to become computationally infeasible or highly manipulable. Observation 1 For k = 1, AV , SAV , PAV , and MAV coincide. Justified Representation We first define a minimal requirement that any reasonable approval based voting rule should satisfy: unanimity prescribes that if there exists a set of candidates that are approved by all the agents, then at least one of these candidates is selected. It is easy to identify a candidate that is unanimously approved. Hence unanimity is computationally easy to achieve. Note that for k = 1, our definition of unanimity coincides with definition of unanimity used for social choice functions: if agents have strict preferences and maximally prefer a certain alternative, then that particular alternative should be selected. We now introduce the main concept discussed in the paper. Setup and voting rules We consider the social choice setting (N, C) where N = {1, . . . , n} is the set of agents and C = {c1 , . . . , cm } is the set of candidates. Each agent expresses an approval ballot Ai ⊂ C that represents the subset of candidates that he approves of, yielding a set of approval ballots A = {A1 , . . . , An }. We will consider approval-based multiwinner rules that take as input (N, C, A, k) and return the subset W ⊆ C of size k that is the winning set (Marshall and Kilgour 2012). 1 Definition 1 (Justified representation) An approvalbased rule satisfies justified representation if there does not exist a profile of ballots and a set N ∗ of at least d|N |/ke T agents such that i∈N ∗ Ai 6= ∅, and no candidate in S i∈N ∗ Ai is selected in the winning set. Approval Voting (AV ) AV finds a set W P ⊆ C of size k that maximizes the total score App(W ) = i∈N |W ∩ Ai |. AV has been adopted by several academic and professional societies such as the Institute of Electrical and Electronics Engineers (IEEE) and the International Joint Conference on Artificial Intelligence. The idea is that if k candidates are to be selected, then set N ∗ that is completely unrepresented can demand that at least one of their unanimously approved candidates should be selected. We say that N ∗ is a violation of justified representation. Satisfaction Approval Voting (SAV ) An agent’s satisfaction is the fraction of his or her approved candidates that are elected. SAV maximizes the sum of such scores. Formally, SAV finds a set W ⊆ C of size k that maximizes P i| Sat(W ) = i∈N |W|A∩A . The rule was proposed with the i| aim of representing more diverse interests than AV (Brams and Kilgour 2010). Theorem 1 Justified representation =⇒ unanimity. Proof: We show that a rule that does not satisfy unanimity does not satisfy justified representation. Assume that a rule is not unanimous. Then, all the agents approve of a certain candidate but no such candidate is selected. Since ∗ 0 |N T | ≥ |N |/k, there exists a set N = N such that C = A = 6 ∅, and no candidate that has been approved by i∈N ∗ i any agent in N ∗ has been selected. Thus justified representation is violated. 2 Proportional Approval Voting (PAV ) In PAV , an agent’s satisfaction score is 1 + 1/2 + 1/3 + · · · + 1/j where j is the number of his or her approved candidates that are selected in W . Formally, PAV finds a set W P ⊆ C of size k that maximizes the total score PAV (W ) = i∈N r(|W ∩ Pp Ai |) where r(p) = j=1 1j . It has recently been shown that computing PAV is NP-hard (Aziz et al. 2014). Since justified representation is based on a highly natural idea of proportionality, it bears similarities to the Droop proportionality criterion and different variants of the solid coalition property (Dummett 1984; Tideman and Richardson 2000; Elkind et al. 2014). However, these proportionality notions are designed for elections for strict preferences and focus on the election of certain candidate(s) which have sufficient support. Minimax Approval Voting (MAV ) MAV (minimax approval voting) returns a set W that minimizes the maximum distance between W and approvals of agents. The distance d(Q, T ) between any two sets Q and T is defined Justified representation of Approval-based Rules 1 If more than one outcome is possible, then ties are broken according to some preset tie-breaking over the candidates. We first observe that all the rules considered in the paper satisfy unanimity. 21 Theorem 2 AV , SAV , PAV , RAV , and MAV satisfy unanimity. Quest for rules satisfying justified representation Although they satisfy unanimity, some of the prominent approval-based multi-winner rules do not satisfy justified representation. We wonder whether any of rules discussed so far that do not or may not satisfy justified representation can be modified so as to satisfy justified representation. We propose adaptations of MAV and PAV . Theorem 3 AV and SAV do not satisfy justified representation. • Modified Proportional Approval ). MaxP Voting (MPAV 0 imize score mpav(C 0 ) = r(|A ∩ A |) where i i∈N Pj 1 r(1) = k + 1 and r(j) = k + 1 + i=1 kj−2 for j ≥ 2. Proof: Consider the following profile and let k = 3. A1 A3 A5 A7 A8 A9 : : : : : : {a, b, d} {a, b, d} {a, b, d} {c, x1 , x2 , x3 , x4 , x5 , x6 , x7 } {c, x1 , x2 , x3 , x4 , x5 , x6 , x7 } {c, x1 , x2 , x3 , x4 , x5 , x6 , x7 } A2 : {a, b, d} A4 : {a, b, d} A6 : {a, b, d} • k-Minimax (k-MAV ). Requires agents to only approve k candidates and then use MAV . Theorem 5 MPAV satisfies justified representation. Proof Sketch: Let us consider a situation in which MPAV violates justified representation. This happensTwhen there exists a S set N ∗ such that |N ∗ | ≥ |N |/k, C 0 = i∈N ∗ Ai 6= ∅, and i∈N ∗ Ai does not intersect with the winning set. We know that |N ∗ | = c ≥ n/k. Then |N \ N ∗ | = n − c ≤ ck −c = c(k −1). Now let us look at the P winning set B. The ∗ total MPAV score of B is as follows: i∈N \N ∗ r(|N ∩ Ai |). We need to place the candidate in positions for each agent in N \ N ∗ in a way so as to maximize the contributed score of the candidates that contributed the least total score. Such a score is maximized if each candidate in the winning is approved by each agent in N \N ∗ . The total MPAV score of the candidates is less than or equal to Then {a, b, d} is selected by AV as well as SAV . But N ∗ = {7, 8, 9} can demand for one of {c, x1 , x2 , x3 , x4 , x5 , x6 , x7 } to be selected. Hence, justified representation is violated due to N ∗ . 2 Next, we show that even an outcome of MAV may not satisfy justified representation. Theorem 4 An outcome of MAV may not satisfy justified representation. Proof: Consider the following profile and let k = 3. A1 A2 A3 A4 A5 A6 A7 A8 A9 : : : : : : : : : k−1 1(1 − k1 c(k − 1)(k + 1) + c(k − 1) 1 − k1 k < c(k − 1)(k + 1) + c(k − 1) k−1 = c(k − 1)(k + 1) + ck. {a, b, d, x1 , x2 , x3 , x4 } {a, b, d, x4 , x6 , x7 , x8 } {a, b, d, x9 , x10 , x11 , x12 } {a, b, d, x13 , x14 , x15 , x16 } {a, b, d, x17 , x18 , x19 , x20 } {a, b, d, x21 , x22 , x23 , x24 } {c} {c} {c} ) We want to show that the contribution of (k + 1)c of candidate x is greater than the average contribution of candidate in B i.e, (k + 1)c > c(k − 1)(k + 1) + ck k This holds iff Then a, b, d are selected by MAV since the maximum hamming distance of {a, b, c} is 6 where as for {a, b, d}, the maximum hamming distance is 4. But N ∗ = {7, 8, 9} can ask for c to be selected. Hence, justified representation is violated.2 2 k(k + 1)c > c(k − 1)(k + 1) + ck iff k 2 + k > (k + 1)(k − 1) + k iff k 2 > (k + 1)(k − 1) iff k 2 > k 2 − 1. As for PAV , it is not clear whether it satisfies justified representation or not. The counter examples that we constructed for other rules do not work to show that PAV does not satisfy justified representation. We conjecture that PAV does satisfy justified representation. We have shown that the average contributed MPAV T score of candidate in B is less than the contribution x ∈ i∈N ∗ Ai can make. Hence, there exists at least one candidate b that has a contributed MPAV score of less than x which means that the total MPAV score can be increased by removing T such a candidate b and replacing it with candidate x ∈ i∈N ∗ Ai . But this is a contradiction because B maximizes the MPAV total score. 2 2 The example in this proof shows that agents 1 to 6 manage to get {a, b, d} selected by simply approving too many dummy candidates of the type x. This shows that MAV can in some situations be easily manipulated. 22 sets of agents of cardinality at least n/k, justified representation can be checked in linear time. Corollary 1 There always exists an outcome that satisfies representation. Next we turn our attention to k-MAV . Theorem 7 It can be checked in linear time whether a committee satisfies justified representation or not. Lemma 1 For k-MAV , if there is a solution in which each agent has one approved candidate then k-MAV will select such a solution. 0 Proof: We Sidentify the set N ⊂ NSof unrepresented agents. Consider i∈N 0 Ai . For each a ∈ i∈N 0 Ai , check whether there exist at least n/k agents in N 0 that approve of a. If there exists such an a, then the committee does not satisfy justified representation. 2 Proof: Such a solution will lead to a maxmin score of at most 2k − 1 whereas for a solution in which an agent is not represented at all has a maxmin score of 2k. 2 Next, we show that a committee that satisfies justified representation can be computed in linear time as well via a greedy algorithm as follows. We refer to the greedy algorithm as GRAV (greedy representation approval voting). Theorem 6 At least one solution of k-MAV satisfies justified representation. Proof Sketch: The maximum possible Hamming distance of an agent’s ballot from the solution can be 2k (when no approved candidates of an agent get selected). If the maxmin score of an MAV solution is less than 2k, then we know that each agent has one approved candidate as a winner. Hence, justified representation is trivially satisfied by a solution for which the maximum Hamming distance is less than 2k. Now let us consider the case when even the k-MAV solution has a maxmin value of 2k. Thus by Lemma 1, there exists no solution for which each agent is represented. We show that the violations can be removed. The distance of 2k remains constant since it is upper-bounded by 2k. Assume that there exists an S a violation of justified representation. T |S| = c ≥ n/k. Thus a candidate x ∈ i∈N ∗ Ai represents c agents. Now let us compute the average number of agents a candidate b ∈ B represents. The maximum number of agents that B can represent is n − c. Therefore the average number of agents represented by candidates in B is less than or equal to (n − c)/k. Hence, there exists at least one candidate b whose removal will lead to less than (n − c)/k agents in N \ N ∗ not being represented anymore. But then at least c ≥ n/k additional agents can be represented by replacing b with x. Thus, whenever there is a violation of justified representation, it can be removed and the total number of agents who are represented by at least one candidate in the winning set increases. The process of resolving violations by swapping b with x can involve n/k iterations. After this we know that each agent is represented. But in this case, justified representation is satisfied by definition since there exists no unrepresented coalition S with a justified complaint. 2 Input: N, C, (A1 , . . . , An ), k. Output: Committee S 0 1 k ←− 0 0 0 0 2 S ←− ∅; N ←− N ; C ←− C; V ←− V ; 0 3 while k < k do 4 Find a candidate c with the highest approval score with respect to (N 0 , C 0 , V 0 ) and put it in S 5 k 0 ←− k 0 + 1 6 Remove all agents from N 0 who approve of c and remove c from V 0 7 if N 0 = ∅ then 8 N 0 ←− N 9 return S Algorithm 1: GRAV Theorem 8 There exists a linear-time algorithm GRAV to compute a representative committee. Proof: Let N ∗ be the final assignment of variable set N 0 under GRAV . We claim that the S satisfies justified representation. Assume for contradiction that S does not satisfy justified representation. This means there exists a candidate b∈ / S that is approved by at least n/k agents in N ∗ who are not represented in S. Since b is unrepresented, this also implies that N 0 was never reset to N . Observe that each time a candidate a0 is placed in S, it has approval score (restricted to the preference profile when it was removed) that is at most the approval score (restricted to the preference profile when a was removed) of any candidate a that was selected in S before a0 . Since b that was not even selected in S is approved by at least n/k agents in N ∗ , this means that each candidate in S has approval score at least n/k in the preference profile when it was removed. Hence, each time a candidate a is removed from C 0 , there are also at least n/k agents removed from N 0 . But this implies that N ∗ is empty and each agent is represented. 2 The result for k-MAV is somewhat unsatisfactory because agents are forced to approve exactly k candidates even if they actually approve of less or more than k candidates. Complexity of Justified representation We note that k-MAV and MPAV , the two rules that we showed satisfy justified representation, are both NP-hard to compute. The fact that MPAV outcome is NP-hard to compute follows from the proof of NP-hardness for PAV in (Aziz et al. 2014). This raises the question whether there exists a rule that is polynomial-time and satisfies justified representation. We first point out that although the naive way of checking justified representation requires going through all Observation 2 For k = 1, GRAV is equivalent to AV . GRAV also has some pitfalls. If a voter is sure that a candidate a will be selected because it has a high enough 23 Example 9 Consider the following profile and let k = 2. approval score, it may have an incentive not to approve of it so that it can affect the selection of some candidate after a. Consider the following profile and let k = 2. A1 : {a, b} A3 : {a} A1 : {a, c} A3 : {a, b} A5 : {a} A2 : {c} A4 : {a} A2 : {b, c} A4 : {a, b} A6 : {a} Then {a, b} and {a, c} are two committees that satisfy representation. However agent 1 can ensure {a, c} is selected by reporting only c has his approved candidate. However, if the rule only satisfies justified representation, agent 1 cannot ensure the selection of {a, c}. Then GRAV returns {a, b} as the outcome. However if agent 1 reports only c as his approved candidate, then by neutrality {a, b} and {a, c} have the same chance of being selected. If ties are broken in favour of b over c, then agent 1 certainly has an incentive to only approve b. Finally, we note that as long as each agent has an approved candidate elected, the resultant committee also satisfies justified representation. However, if no such committee exists, then justified representation can still help refine which committee to select. Previous Approaches to Approval-based Representation Constrained approval voting Although voting rules such as PAV have proportional representation as one of the goals (as implied by the name), some other approaches in approval-based voting were explicitly designed for better ‘representation’. However, the approaches are computationally demanding. Justified representation has more bite than unanimity but is still computational easy to satisfy. There is another approval-based method that concerns a different setting in which there are lower bound and upper bound constraints on the number of candidates that can be elected from different categories of candidates and approval voting is implemented while taking into account these quota constraints for categories (Brams and Pothoff 1990; Straszak et al. 1993). Implementing such constraints again results in NP-hardness of computing such solutions. Moreover, we are interested in the the outcome being a function of the ballots and not on other additional side constraints. Proportional Representation and Approval ballots Proportional representation has been proposed by Monroe (1995) in which agents should be partitioned into k equally-sized groups, and each group assigned to a candidate so as to minimize dissatisfaction of agents from their associated candidates. Chamberlin and Courant (1983) used a similar approach but did not impose the condition of equal sized groups and compensated it by attributing weights to groups of different sizes. Unfortunately, even with approval ballots, finding an optimal representative committee is NPhard in both the Monroe as well as the model by Chamberlin and Courant (Procaccia, Rosenschein, and Zohar 2008). Discussion Unanimity Justified representation + + + + + + + + + ? + + + + + Rule AV SAV MAV PAV k-MAV MPAV GRAV U JRAV EJRAV Threshold approaches Kilgour (2010) mentioned another kind of (threshold) representation which if possible will return a winning set in which for each agent, at least one approved candidate is selected. We note this version of representation is overly strong and unreasonable. If k = 2 and 99 agents prefer the committee {a, b} and the only one agent 100 approves of {c, d}, then threshold representation requires either c or d to be selected. In this sense, threshold representation is overly egalitarian and ignores the relative numbers of agents supporting different candidates. Threshold-based representation gives too much flexibility to agents to manipulate an outcome by only approving their most preferred candidate that has not been selected if they are sure that other approved candidates will be selected in any case. On the other hand, if a rule only satisfies justified representation, then an agent can only get a particular candidate selected if he has sufficient support from other agents. Table 1: Satisfaction of justified representation by approvalbased voting rules. In this paper, we initiated the axiomatic analysis of an important fairness aspect of approval-based multi-winner rules. We identified a useful axiom called justified representation that is a generalization of unanimity. Justified representation is also computationally easier to achieve than trying to represent each agent. It will be interesting to check whether other approvalbased rules satisfy justified representation. A promising rule is PAV (Kilgour 2010). Other rules based on approval may also be worth considering (Betzler, Slinko, and Uhlmann 2013; Skowron, Faliszewski, and Slinko 2013). 24 Using justified representation to devise new rules Justified representation can also be used to formulate new approval-based rules. We mention a couple of rules that seem particularly attractive: Conitzer, V. 2010. Making decisions based on the preferences of multiple agents. Communications of the ACM 53(3):84–94. Dummett, M. 1984. Voting Procedures. Oxford University Press. Elkind, E.; Faliszewski, P.; Skowron, P.; and Slinko, A. 2014. Properties of multiwinner voting rules. In Proceedings of the 13th International Conference on Autonomous Agents and Multi-Agent Systems (AAMAS), 53–60. Elkind, E.; Lang, J.; and Saffidine, A. 2011. Choosing collectively optimal sets of alternatives based on the condorcet criterion. In IJCAI, 186–191. Kilgour, D. M. 2010. Approval balloting for multi-winner elections. In Handbook on Approval Voting. Springer. chapter 6. Laslier, J.-F., and Sanver, M. R., eds. 2010. Handbook on Approval Voting. Studies in Choice and Welfare. SpringerVerlag. LeGrand, R.; Markakis, E.; and Mehta, A. 2007. Some results on approximating the minimax solution in approval voting. In Proceedings of the 6th International Conference on Autonomous Agents and Multi-Agent Systems (AAMAS), 1193–1195. Marshall, E., and Kilgour, D. M. 2012. Approval balloting for fixed-size committees. In Electoral Systems, Studies in Choice and Welfare. Springer. chapter 12, 305–326. Meir, R.; Procaccia, A. D.; and Rosenschein, J. S. 2008. A broader picture of the complexity of strategic behavior in multi-winner elections. In Proceedings of the 7th International Conference on Autonomous Agents and Multi-Agent Systems (AAMAS), 991–998. Monroe, B. L. 1995. Fully proportional representation. The American Political Science Review 89(4):925–940. Procaccia, A. D.; Rosenschein, J. S.; and Zohar, A. 2008. On the complexity of achieving proportional representation. Social Choice and Welfare 30:353–362. Skowron, P. K.; Faliszewski, P.; and Slinko, A. 2013. Fully proportional representation as resource allocation: Approximability results. In Proceedings of the 23rd International Joint Conference on Artificial Intelligence (IJCAI). Straszak, A.; Libura, M.; Sikorski, J.; and Wagner, D. 1993. Computer-assisted constrained approval voting. Group Decision and Negotiation 2(4):375–385. Tideman, N., and Richardson, D. 2000. Better voting methods through technology: The refinement-manageability trade-off in the single transferable vote. Public Choice 103(1-2):13–34. • Utilitarian justified representation (UJRAV) rule: the rule that among all committees that satisfy justified representation, returns a committee that maximizes the sum of approvals. • Egalitarian justified representation (EJRAV) rule: the rule that among all committees that satisfy justified representation, returns a committee that maximizes the the number of representatives of the agent who has the least number of representatives in the winning committee. The computational complexity of winner determination for these rules is an interesting problem. Finally, analyzing the compatibility of justified representation with other important properties such as strategyproofness for dichotomous preferences is another avenue of future research. Acknowledgements The authors acknowledge the helpful comments of Steven Brams. This material is based upon work supported by the Australian Government’s Department of Broadband, Communications and the Digital Economy, the Australian Research Council, the Asian Office of Aerospace Research and Development through grant AOARD-12405. References Aziz, H.; Gaspers, S.; Gudmundsson, J.; Mackenzie, S.; Mattei, N.; and Walsh, T. 2014. Computational aspects of multi-winner approval voting. In Proceedings of the 8th Multidisciplinary Workshop on Advances in Preference Handling. Betzler, N.; Slinko, A.; and Uhlmann, J. 2013. On the computation of fully proportional representation. Journal of Artificial Intelligence Research 47:475–519. Brams, S. J., and Fishburn, P. C. 2007. Approval Voting. Springer-Verlag, 2nd edition. Brams, S. J., and Kilgour, D. M. 2010. Satisfaction approval voting. Technical Report 1608051, SSRN. Brams, S. J., and Pothoff, R. F. 1990. Constrained approval voting: A voting system to elect a governing board. Interfaces 20(2):67–80. Brams, S. J.; Kilgour, D. M.; and Sanver, M. R. 2006. Mathematics and democracy: recent advances in voting systems and collective choice. In How to elect a representative committee using approval balloting. Springer. 83–96. Brams, S. J.; Kilgour, D. M.; and Sanver, M. R. 2007. A minimax procedure for electing committees. Public Choice 132:401–420. Chamberlin, J. R., and Courant, P. N. 1983. Representative deliberations and representative decisions: Proportional representation and the borda rule. The American Political Science Review 77(3):718–733. 25