COLLECTIVE DECISION MAKING Pierre Dehez CORE University of Louvain pierre.dehez@uclouvain.be Outline 1. Preferences, utility and choices 2. Cardinal welfarism distributive justice utilitarism vs egalitarism Nash bargaining social welfare orderings transferable utility games 3. Ordinal welfarism the case of two alternatives social choice procedures impossibility theorems possibility theorems 2 References Austen-Smith D. and J. Banks, Positive political theory I: Collective preferences, University of Michigan Press, 1999. Austen-Smith D. and J. Banks, Positive political theory II: Strategy and structure, University of Michigan Press, 2005. Brams S., Game theory and politics, Dover, 2004. Brams S., Mathematics and democracy, Princeton University Press, 2008. Moulin H., Axioms of cooperative decision making, Cambridge University Press, 1998.* Moulin H., Fair division and collective welfare, MIT Press, 2003.* Taylor A., Mathematics and politics, Springer-Verlag, 1995. Peyton Young H., Equity. In theory and Practice, Princeton University Press, 1995. Handbook of social choice and welfare, Elsevier, 2002. * Moulin's monographies have inspired some of the material presented here. 3 1. Preferences, utility and choices 4 Preferences Preferences over a set A of alternatives are defined by a (binary) relation over A: a b b is not preferred to a from which the strict preference and indifference relations are deduced: a b a is preferred to b [a a b indifference between a and b [a b and b a] b and b a] 5 Preferences are rational if they verify the following properties: - completeness: [a b or b a] - reflexivity: a a for all a A - transitivity: [a b and b for all a, b A c] a c Completeness is by far the most demanding assumption! A relation satisfying reflexivity and transitivity is a preorder. We denote by L(A) the set of preorders on a set A. 6 Ordinal utilities A preference preorder carries no information on the intensity of preferences: if a is preferred to b and c is preferred to d, we don't know whether a is "more preferred" to b than c is preferred to d Under minimal assumptions, preferences can be represented by a utility function u : A : a A u (a) which associates a real number to each alternative, such that: u(a) u(b) a b 7 As such, this is an ordinal representation of preferences: only the sign of the difference u(a) u(b) provides an information on the preferences between a and b u (a) u (b) 0 a b u (a) u (b) 0 a b As a consequence, u and v T (u ) where T : is an arbitrary increasing transformation, both represent the same preferences. T(u) = u3 is the simplest nonlinear transformation with range . 8 Choices Given a set of alternatives A and preferences element in A: L( A), a choice is a best a* A a* a for all a A or a * maximizes u(a) on A There may be several best elements. The set of solutions is called the choice set. 9 Let C ( A, ) denote the choice set associated to a set A of alternatives and a preference relation . Then: a, b C ( A, ) a b There is indifference between the elements of a choice set. In the multivalued case, a neutral mechanism is necessary to eventually retain a unique alternative. For instance a random mechanism. 10 Cardinal utilities Utilities are cardinal if utility difference have a meaning: u(a) u(b) u(c) u(d ) 0 means that a is preferred to b more intensely than c is preferred to d. Cardinal utility function are defined up to an increasing affine transformation: u and v au b, a 0 are utility functions representing the same preferences. 11 Collectivity: preference profiles Consider a set A of alternatives and n individuals indexed by i running from 1 to n, each having a preference relation i L( A). A preference profile P specifies a preference relation for each member of the group: P ( 1,..., n ) L ( A) n A utility profile can be associated to any alternative a A: u(a) u1 (a),..., un (a) n 12 One of the questions addressed by social choice is the determination of a collective preference ordering for comparing utility profiles. There are several levels of independence that collective preferences may satisfy: 1. Ordinal, non-comparable: full independence 2. Ordinal, comparable: independence of common utility space 3. Cardinal, non-comparable: independence of utility scales 4. Cardinal, partially comparable: independence of zero utilities 5. Cardinal, comparable: independence of utility scales and zero utilities 13 Given a set of alternative A and a preference profile on A ( 1,..., n ) L ( A) n represented by utility functions u1,…,un we define the attainable utility set U ( A) u n u u1 (a ),..., un (a) , a A The problem is then to pick up a point in this set, possibly given the specification of a disagreement point d in U(A). 14 1. Ordinal, non-comparable: full independence This is the situation where each individual utility level is defined up to an arbitrary increasing transformation: (u1 ,..., un ) (v1 ,..., vn ) (T1 (u1 ),..., Tn (un )) (T1 (v1 ),..., Tn (vn )) where the Ti's are arbitrary increasing transformation from into . 15 2. Ordinal, comparable: independence of common utility space This is the situation where individual utility levels are defined up to an arbitrary and common increasing transformation: (u1 ,..., un ) (v1 ,..., vn ) (T (u1 ),..., T (un )) (T (v1 ),..., T (vn )) where T is an arbitrary increasing transformation from into . 16 3. Cardinal, non-comparable: independence of utility scales This is the situation where each individual utility level is defined up to an increasing and affine transformation: (u1 ,..., un ) (v1 ,..., vn ) (a1u1 b1 ,..., anun bn ) (a1v1 b1 ,..., an vn bn ) for all ai , bi , ai 0. 17 4. Cardinal, partially comparable: independence of zero utilities This is the situation where each individual utility level is defined up to an increasing and affine transformation: (u1 ,..., un ) (v1 ,..., vn ) (u1 b1 ,..., un bn ) (v1 b1 ,..., vn bn ) for all b1 ,..., bn . Alternatively: (u1 ,..., un ) (v1 ,..., vn ) (u1 v1,..., un vn ) (0,...,0) 18 5. Cardinal, comparable: independence of utility scales and zero utilities This is the situation where individual utility levels are defined up to an increasing and affine common transformation: (u1 ,..., un ) (v1 ,..., vn ) (a u1 b,..., a un b) (a v1 b,..., a vn b) for all a, b , a 0. 19 2. Cardinal welfarism 20 2.1 Distributive justice 21 "Equal treatment of equals" is the basic principle of distributive justice. It is a minimal and clear requirement of fairness. "Unequal treatment of unequals" instead is a vague principle. "Equals should be treated equally and unequals unequally, in proportion to the relevant similarities and differences" (Aristoste) 22 Liberalism: the social order emerges from the interaction of free wills. Methodological individualism is at the root of liberalism. Individuals are characterized by values, rights and obligations. Distributive justice has two sides: - procedural justice: is the distribution of rights fair ? - end-state justice: is the outcome fair ? We start with a simple problem of sharing a resource. 23 We assume that a utility index can be associated to each individual: vi ui ( xi ) where xi denotes the share of individual i in the resource. It is a cardinal utility and utility levels can be compared. Its definition depends upon the context. It is an "objective" index and the individual is not responsible for its shape. It is the information that a benevolent dictator needs to decide on the allocation of resources in a particular context. 24 Principle 1: ex ante equality There are basic rights like freedom of speech, access to education, freedom of religion, equal political rights (one person, one vote),… They induce ex ante equality: equal claim to the basic resources. Private ownership or differences in status (for instance seniority) are instances of unequal exogenous rights which justify unequal treatment. 25 Principle 2: ex post equality … justifies unequal shares of resources to compensate for involuntary differences in individuals' primary characteristics like nutritional needs, health,… If ui is an objective utility index for individual i resulting from his/her primary characteristics, this principle allows for or equivalently ui ( x) u j ( x) ui ( xi ) u j ( x j ) xi x j This principle amounts to equalization of utilities. 26 Principle 3: reward or penalize … justifies unequal shares yi's of resources to compensate for voluntary differences in individuals' characteristics: - past sacrifies justify a larger share - past abuses justify a lesser share How to reward individual contributions ? The answer if difficult when there are externalities (extraction of exhaustible resources, division of joint costs or surpluses). 27 Principle 4: best use of the resources (fitness) …resources must go to those that can make the best use them. Fitness justifies unequal treatment by differences in talent, independently of basic rights, needs or merits. Two definitions: sum-fitness: maximization of the sum of the individual utilities efficiency-fitness: Pareto optimality Sum-fitness implies efficiency fitness. 28 How should the benevolent dictator use these four partially conflicting principles very much depends upon the context. Examples: - access to the lifeboat, - allocation of organs for transplant, - seat rationing, - political rights. 29 Lifeboat exogenous rights: strict equality (lottery) or priority ranking based on social status or wealth compensation: priority to the weak ones (equality of ex-post survival chances) reward: exclude those responsible for the sinking ship… fitness: keep the crew, the women, the children,… 30 Transplants exogeneous rights: strict equality (lottery) or priority ranking based on social status or wealth compensation: priority to those suffering most or whose life expectancy is the shortest reward: priority to seniority on the waiting list fitness: maximization of the chances of success 31 Seats: auctioning or queuing exogenous rights: only a lottery would induce a strict equality reward: queuing reward efforts while auctioning does not fitness: queuing meets sum-fitness but involves a waste of time auctioning is better if individuals are comparable, because otherwise it favors the rich 32 Political rights fitness and reward: justify unequal voting rights which were commonplace in the past exogenous rights: justifies equal rights (beyond some obvious limitations justified by fitness) compensation: there are many examples of situations where voting rights are not equal (EU distribution of votes among countries supposed to take into differences in population sizes) 33 Allocation methods A given amount of some commodity has to be divided between a given number of individuals and each individual has a claim. The commodity could be a "good" or a "bad": - for a good, individuals express demands - for a bad, individuals have liabilities There may be an excess or a deficit. 34 Data: a set N = {1,…n} of "players" an amount E > 0 to be allocated players' claims: d1,…,dn > 0 Problem: find an allocation x = (x1,…,xn) sucth that x(N) = E. Two cases: deficit: d ( N ) E surplus: d ( N ) E Notation: for all S N : x( S ) xi iS 35 Examples: joint venture: E is the revenue generated by the cooperation and the di's are the stand-alone revenues (surplus) bankcruptcy: E is the firm's liquidation value and the di's are the creditors' claims (deficit) inheritance: E is the value of the deceased's estate and the di's are the heirs' deeds (deficit or surplus) taxation: E is the tax to be levied and the di's are the taxable incomes (deficit) 36 Assumption: equal exogenous rights the allocation depends only on the distribution of claims or liabilities An allocation method is a rule that associates an allocation ( x1 ,..., xn ) ( E, d1,..., dn ) to any given allocation problem ( E, d1 ,..., dn ) such that x( N ) E. 37 Proportional rule (in the case of a surplus or a deficit) di xi E d (N ) satisfies: xi di for all i in case of a surplus xi di for all i in case of a deficit and E xi x j (di d j ) for all i and j d (N ) 38 x2 x1 d1 E d1 d 2 x2 d2 E d1 d 2 x1 x2 E x2 PROP d2 0 d2 x1 d1 d1 x1 39 Equal surplus rule (in case of a surplus) 1 xi di ( E d ( N )) n satisfies E d (N ) 0 xi di for all i xi x j di d j for all i and j 40 x2 SURPLUS: d1+d2 < E and d1 > d2 x2 x1 x2 x1 (d2 d1 ) E d1 d 2 2 E d 2 d1 x2 2 x1 ES d2 0 d1–d2 d1 x1 41 x2 SURPLUS: d1+d2 < E and d1 < d2 x2 x1 (d2 d1 ) x2 x1 ES E d1 d 2 2 E d 2 d1 x2 2 d2 x1 d2–d1 0 d1 x1 42 Uniform gain rule (in case of a surplus) xi Max ( z, di ) n where z satisfies E Max ( z, di ) i 1 satisfies xi di for all i. Here z can be interpreted as the common gain. This rule is also called "constrained" egalitarian. 43 y 2z y = f(z) y d1 z d1 d2 2d1 f ( z) Max( z, d1 ) Max( z, d2 ) d1+d2 d1 0 d2 d1 z 44 x2 x2 x1 UG d2 0 d1 x1 45 x2 SURPLUS: d1+d2 < E x2 x1 UG x2 x1 (d2 d1 ) ES PROP x2 d2 x1 d1 d2 0 d1–d2 d1 x1 46 Uniform gain rule (in case of a deficit) xi Min ( z, di ) n where z satisfies E Min ( z, di ) i 1 satisfies xi di for all i. Here z can be interpreted as the common gain. This rule is also called "constrained" egalitarian. 47 y 2z y = f(z) y d2 z d1+d2 d1 d2 2d2 f ( z) Min( z, d1 ) Min( z, d2 ) d2 0 d2 d1 z 48 x2 DEFICIT: d1+d2 > E x2 x1 d2 UG 0 d1 x1 49 Uniform loss rule (in case of a deficit) xi Max (di z, 0) or di xi Min( z, di ) n where z satisfies E Max (di z , 0) i 1 satisfies xi di for all i. The idea is to substract the same amount from the claims subject to the non-negativity constraint: z is the common loss (ex post deficits are equalized). This rule is also called levelling. 50 y = f(z) DEFICIT: d1+d2 > E d1+d2 d1 d2 d1 f ( z) Max (d1 z,0) Max (d2 z,0) d1–d2 0 d2 y d1 d2 2 z d1 z y d1 z 51 x2 DEFICIT: d1+d2 > E x2 x1 x2 x1 (d2 d1 ) E d1 d 2 2 E d 2 d1 x2 2 x1 d2 UL 0 d1–d2 d1 x1 52 x2 DEFICIT: d1+d2 > E d2 UG PROP UL 0 d1–d2 d1 x1 53 x2 UG ES PROP d2 UG PROP UL 0 d1–d2 d1 x1 54 Proportional (surplus/deficit) Equal surplus (surplus) Uniform gain (surplus) Uniform gain (deficit) Uniform loss (deficit) xi E di d (N ) xi di 1 ( E d ( N )) n xi Max ( z , d i ) where z is such that xi E xi Min ( z, di ) where z is such that xi Max (di z, 0) where z is such that x E x E i i 55 Algorithm for computing the uniform gain solution - divide E in equal parts xi Max ( z, di ) xi Min ( z, di ) - identify the individuals with the "wrong" share: E i such that di in case of a deficit n xi di E i such that di in case of a surplus n - reduce E accordingly and repeat the procedure with the remaining individuals… 56 Deficit: n = 5, E = 40 and d = (20, 16, 10, 8, 6) d ( N ) 60 E E 8 x4 8 and x5 6 5 E 14 3 8.7 x1 x2 x3 8.7 57 Surplus: n = 5, E = 80 and d = (20, 16, 10, 8, 6) d ( N ) 60 E E 16 x1 20 and x2 16 5 E 36 3 14.7 x3 x4 x5 14.7 58 Algorithm for computing the uniform loss solution xi Max (di z, 0) - apply the equal surplus solution - identify the individuals with a negative share: 1 i such that di ( E d ( N )) 0 n xi 0 - repeat the procedure with the remaining individuals… 59 Deficit: n = 5, E = 20 and d = (20, 16, 10, 8, 6) 1 zi di (40) 5 z (12, 8, 2, 0, 2) x4 x5 0 1 zi di (26) 3 z (11.3, 10.7, 1.3) x (11.3, 10.7, 1.3, 0, 0) 60 Deficit: n = 5, E = 50 and d = (20, 16, 10, 8, 6) 1 zi di (10) 5 z (18,14,8, 6, 4) x (18,14,8, 6, 4) 61 deficit E= di = 20 16 10 8 6 20 PRO 6.7 5.3 3.3 2.7 2 UG 4 4 4 4 4 UL 11.3 7.3 1.3 0 0 PRO 13.3 10.7 6.7 5.3 4 UG 8.7 8.7 8.7 8 6 UL 16 12 6 4 2 PRO 16.7 13.3 8.3 6.7 5 UG 13 13 10 8 6 UL 18 14 8 6 4 PRO 26.7 21.3 13.3 10.7 8 UG 20 16 14.7 14.7 14.7 ES 24 20 14 12 10 PRO 40 32 20 16 12 UG 24 24 24 24 24 ES 32 28 22 20 18 40 50 surplus 80 120 d i 60 62 Bankcruptcy: deficit If creditors have equal exogenous rights, it is the proportional solution that emerges. In reality, there are priorities that may be implied by exogenous rights. Medical supplies: deficit (di stands for the need of patient i) The proportional solution is hardly acceptable in this context. The uniform loss solution imposes itself if reducing the quantity of the drug is equally bad for all (ex: insuline). The uniform gain solution is appropriate if the drug is not essential (ex: sleeping pills). 63 Fund raising: surplus (di stands for the contribution of donor i) The proportional solution is definitely unfair: it penalizes the generous donors. Nor is the equal surplus solution. Uniform gain is the most naturel solution: it requires the less generous donors to contribute first. Fund raising: deficit Uniform loss is not acceptable because it gives a uniform rebate irrespectively of the contributions. The proportional solution is definitely more appropriate. 64 The uniform gain may be acceptable as well. Indeed, it can be computed through the following alternative algorithm, assuming that claims are ordered as follows: d1 d2 ... dn - decrease 1's claim by (d1 – d2) x = (d2, d2, d3,… dn) - decrease 1 and 2's claims by (d2 – d3) x = (d3, d3, d3, d4, … dn) - decrease 1, 2 and 3's claims by (d3 – d4) x = (d4, d4, d4, d4, … dn) … until arriving below E. The difference E isthen xi returned uniformly to the players whose claims have been decreased. 65 Deficit: n = 5, E = 40 and d = (20, 16, 10, 8, 6) - decrease 1's claim by (d1 – d2) = 4 x = (16, 16, 10, 8, 6) - decrease 1 and 2's claims by (d2 – d3) = 6 x = (10, 10, 10, 8, 6) - decrease 1, 2 and 3's claims by (d3 – d4) = 2 x = (8, 8, 8, 8, 6) The total is then 38. So add 2/3 to the first three players: x (8.7, 8.7, 8.7, 8, 6) 66 Four desirable properties 1. invariance with respect to transfers 2. truncation property 3. concession property 4. consistency 67 1. Invariance to transfers if i and j "merge" into a single individual, is the resulting share equal to the sum of the individuals' shares ? Only the proportional rule is invariant to transfers. The uniform gain rule is not: merging leads to a smaller or equal share. The uniform loss rule is not: merging leads to a higher or equal share. 68 2. Truncation property In case of a deficit, a solution satisfies the truncation property if truncating the claims to E di di Min di , E does not affect the resulting allocation. The uniform gain rule satisfies the truncation property. The uniform loss rule and the proportional rule do not. 69 3. Concession property In case of a deficit, we define the concession of N\i to individual i by: zi Max 0, E d ( N \ i) Given an allocation rule, consider the following 2-step procedure: - allocate zi to individual i - apply the allocation rule to the problem of dividing what remains E E zi according to the reduced claims di di zi . 70 An allocation rule has the concession property if this 2-step procedure reaches the same allocation. The uniform loss rule satisfies the concession property. The uniform gain rule and the proportional rule do not. 71 4. Consistency An allocation rule is consistent if for all problem (E,d) and all subsets S in N: x (E, d ) x where x S S ( x(S ), d S ) ( xi | i S ). Pairwise consistency requires that condition to hold for any pair of individuals. For continuous and symmetric rules, pairwise consistency implies consistency. 72 Example: n = 5, E = 50 and d = (20,16,10,8,6). The uniform gain solution is x = (13,13,10,8,6). Looking at S = {1,2,3} and applying the solution to the problem defined by E = 36 and d = (20,16,10), we get x = (13,13,10). The uniform loss solution is x = (18,14,8,6,4). Looking at S = {1,2,3} and applying the solution to the problem defined by E = 40 and d = (20,16,10), we get x = (18,14,8). 73 Application: taxation Here E = T is the tax to be levied and di = yi represents i's taxable income. It is assumed that we are in the deficit case: y( N ) T . A taxation method is a function which associates taxes t (T , y ) to any taxation problem (T,y) such that t i T and 0 ti yi for all i There are three classical taxation methods: flat tax, head (or poll) tax and levelling tax. 74 Flat tax is the proportional solution: tj ti ti yi where yi y j yi T Head tax is the uniform gain solution: ti Min ( z, yi ) where z is such that t i T Levelling tax is the uniform loss solution: ti Max ( yi z,0) where z is such that t i T 75 y1 y2 t2 t2 t1 t2 t1 ( y1 y2 ) t2 y2 t1 y1 y2 Head Flat Levelling 0 y1–y2 y1 t1 76 y1 y2 t2 t2 t1 t2 t1 ( y1 y2 ) t2 y2 t1 y1 y2 Head Flat Levelling 0 y1–y2 y1 t1 77 Principles 1. Fair ranking A higher income justifies both a higher tax burden and a higher aftertax income : ti t j yi y j yi ti y j t j Under this principle, equal incomes are taxed equally. 78 y1 y2 Max 0, t1 ( y1 y2 ) t2 Min y2 , t1 t2 t2 t1 ( y1 y2 ) t2 t1 y2 Head fair Levelling 0 y1–y2 y1 t1 79 2. Progressive tax A higher income justifies a higher the tax rate: tj ti ti yi yi y j or yi y j tj yj 3. Regressive tax A higher income justifies a lower the tax rate: tj ti yi y j yi y j 80 t2 regressive region: head tax is the most regressive method T t2 y2 t1 y1 y2 0 y1–y2 T y1 t1 81 t2 progressive region: levelling tax is the most progressive method T y2 t2 y2 t1 y1 0 y1–y2 T y1 t1 82 The exponential method is defined by: ti Min yip , yi for some p 0 where is choosen such that t i T. It is progressive for p > 1 and regressive for p < 1. It is the flat tax for p = 1. It is the head tax for p = 0. 83 t2 y2 0 y1–y2 y1 t1 84 Equal sacrifice "Equality of taxation means equality of sacrifice. It means apportioning the contribution of each person towards the expenses of the government so that he shall feel neither more nor less inconvenience from his share of the payment than every other person experiences from his." John Stuart Mill, Principle of Economics, 1848 85 Let ui(y) be the utility associated to income y by individual i. Equal sacrifice means choosing taxes in such a way that differences of utilities are equalized: ui ( yi ) ui ( yi ti ) z for all i To avoid interpersonal utility comparisons, we postulate a common utility function u (a kind of social norm): u( yi ) u( yi ti ) z for all i Mill proposed to use the Bernoulli utility function log y. 86 u ( y) log y yields the proportional tax: yj tj yi ti yi ti y j t j yi y j Equal relative sacrifice means choosing taxes in such a way that ratios of utilities are equalized: u ( yi ti ) z for all i u ( yi ) It is merely equivalent to equal absolute sacrifice: the log of the ratio equals the difference of the log. 87 Proposition Equal sacrifice implies fair ranking if and only if u is increasing and concave. u ( yi ti ) u ( y j t j ) u ( yi ) u ( y j ) yi y j yi ti y j t j u ( yi ) u ( yi ti ) u ( y j ) u ( y j t j ) yi ti y j t j ti t j 88 Proposition The equal sacrifice method is progressive if and only if the function y.u'(y) is non-increasing in y (i.e. u is more concave than the log function). Proposition The equal sacrifice method is regressive if and only if the function y.u'(y) is non-decreasing in y (i.e. u is less concave than the log function). 89 The contested garment rule "Two people cling to a garment. The decision is that one takes as much as his grasp reaches, the other takes as much as his grasp reaches, and the rest is divided equally among them." (Talmud) Hence the "contested garment rule" for n = 2 is given by: where 1 xi zi E z1 z2 i 1, 2 2 zi Max 0, E d ( N \ i) E Min E, di i 1, 2 90 The solution can then be alternatively written as: x1 1 E Min E , d1 Min E , d 2 2 1 x2 E Min E , d 2 Min E , d1 2 The contested garment rule satisfies both the truncation property and the concession property. Actually, it is the only 2-person rule satisfying these two properties. They define it. An allocation rule has the contested garment property if, when applied to a 2-person problem, it coincides with the contested garment solution. 91 2.2 Egalitarism vs utilitarism 92 Egalitarian vs utilitarian solutions egalitarian solution (compensation): find ( x1,..., xn ) such that ui ( xi ) u j ( x j ) for all i, j and x( N ) E utilitarian solution (sum-fitness): find ( x1,..., xn ) that maximizes u (x ) i i subject to x( N ) E 93 The egalitarian solution solution may not be defined. The proper formulation should instead be the following: find ( x1 ,..., xn ) 0 such that x( N ) E and xi 0 ui ( xi ) Min j u j ( x j ) Whether or not the utility function are concave (decreasing marginal utility) impacts the comparison of the two solutions. In the concave case, the two solutions are in some sense identical. Furthermore, the three solutions studied earlier turns out to be special cases. 94 The link between the two solutions when utility functions are increasing and concave (and differentiable) appears by comparing the revised definition of the egalitarian solution and the first order condition associated to the utilitarian solution: xi 0 ui ( xi ) Min j u j ( x j ) xi 0 ui( xi ) Max j uj ( x j ) Hence, the utilitarian solution with utility functions ui corresponds to the egalitarian solution with utility functions – ui'. If concavity is quite natural in a context of income distribution, convexity may be adequate in other context e.g. medical rationing. 95 The following example illustrates the role of concavity. A quantity E of some resource has to be divided between n individuals. Each individual i is initially endowed with a quantity i and his/her preferences are represented by ui ( xi ) u(i xi ) where u is some common (base) and strictly concave utility function. Mill has shown that the egalitarian and utilitarian solutions coincide: they both equalize the final outcome i xi . 96 The egalitarian solution reads: xi 0 u (i xi ) Min jN u ( j x j ) i xi Min jN ( j x j ) because u is increasing. Hence i xi j x j for all i, j. We observe that this solution is equivalent to the uniform gain solution applied to the problem of dividing the amount E ' E ( N ) with claims di i . 97 The utilitarian solution is the solution of the following maximization problem: Max u ( j x j ) subject to: x j E xi 0 i 1,..., n Using the 1st order conditions, we have: xi 0 u(i xi ) Max jN u( j x j ) i xi Max jN ( j x j ) because u' is decreasing. Hence i xi j x j for all i, j. 98 If now u is a strictly convex function, the egalitarian solution is unchanged. Being a uniform gain solution, it is independent of the choice of the base utility function that only needs to be increasing. The utilitarian solution instead allocates all the resource to the richest individual ! Indeed if positive amounts xi and xj are allocated to the i and j such that i j strict convexity implies: u(i xi ) u( j x j ) u(i xi x j ) u( j ) i.e. transferring xj to i increases the sum of the utilities. 99 Another example Assume the utility functions are of the form ui ( xi ) i u( xi ) where the i's are positive "productivities" and u is some base and strictly concave utility function such that u(0) = 0. Here the two principles give opposite recommendations. 100 The egalitarian solution simply equalizes utilities: iu( xi ) j u( x j ) for all i, j The utility function u being strictly increasing, shares and productivities are negatively correlated: i j xi x j The utilitarian solution is defined by the 1st order conditions iu( xi ) j u( x j ) for all i, j By strict concavity, shares and productivities are now positively correlated: i j xi x j 101 2.3 Nash bargaining 102 Consider a game in strategic (normal) form (S1, S2, u1, u2) involving two players. We denote by A the set of consequences, allowing for correlated strategies and we work directly on the expected utility set U u 2 u u1 (a ), u2 (a ) , a A Players may agree on a choice of strategies, knowing that in case of disagreement, they find themself in some situation that corresponds to a pair of utilities d = (d1,d2) U(A). One possibility is to refer to prudent (MaxMin) strategies in which case di is the security level of player i. 103 A bargaining problem is defined by pair (U,d) where U is a subset of 2 that is closed, convex and bounded above d is a point in U such that there exists some u U, u >> d Example: the battle of sexes strategic form correlated strategies a2 b2 a1 2,1 0,0 b1 0,0 1,2 a2 b2 a1 p1 p2 b1 p3 p4 0 pi 1 and p i 1 104 u2 (1,2) U (0,0) battle of sexes a2 b2 a1 2,1 0,0 b1 0,0 1,2 (2,1) u1 105 In general, U is the convex hull of the utility pairs corresponding to pure strategies: u2 U C7 U C8 u1 106 Bargaining problem need not result from a game situation. This is the case of allocation problems like the bankcruptcy problem. u2 L C1 + C2 > L C2 U d = (0,0) C1 L u1 107 A solution to a bargaining problem (U,d) is a point u* in U satisfying the following minimal properties: collective rationality: u U such that u u * and u u * individual rationality:u* d We look for a rule associating a solution to any bargaining problem (U,d). A bargaining problem (U,d) is symmetric if d1 = d2 and inter-changing the players results in the same set U i.e. the 45° line is a symmetry axis of U. 108 individual + collective rationality u2 d2 d U d1 u1 109 symmetric bargaining problem 45° d 110 Three natural axioms to be imposed on a rule : Efficiency (collective rationality): there is no u U such that u > (U,d) Individual rationality : 1(U,d) d1 and 2(U,d) d2 Symmetry: if (U,d) is symmetric then 1(U,d) = 2(U,d) 111 These three axioms determine the solution of symmetric bargaining problems: (U,d) d 112 u2 battle of sexes (1,2) 3 3 ( , ) 2 2 U (0,0) (2,1) u1 113 Because utilities are expected utilities we need the following further axiom. Independence with respect to preference representation (covariance) (U,d) (V,c) wherevi = ai + bi ui (bi > 0) ci = ai + bi di i(V,c) = ai + bi i(U,d) (i = 1,2) Indeed we are in a cardinal framework with non comparable utilities (independence of utility scales). 114 ui di vi bi di b2=5 (4.5, 4) d2=3 b2-d2=2 (2.5, 1) 1 (0.5, 0.5) 1 d1=2 b1-d1=5 b1=7 115 This additional axiom determines the solution to bargaining problems whose individually rational boundary is a line segment: (U,d) d It is indeed just the middle of that segment. 116 The following axiom extends the solution to all bargaining problems. Independence with respect to irrelevant alternatives: If (U,d) and (V,d) are such that U V and (V,d) U then (U,d) = (V,d). 117 (V,d ) U (U,d ) = (V,d ) UV 118 Here the bargaining problem (V,d) is defined by the line tangent to U at its mid point: u2 this is the Nash solution (U,d) U d u1 119 contour curves are rectangular hyperbolas u2 u* (u1 – d1)(u2 – d2) = constant U d u1 120 Indeed the line segment tangent to a rectangular hyperbola and restricted to the axis is divided in its middle at the tangency point. Hence, the Nash solution is nothing but the solution of the maximization of the product of the gains on U: MaxuU (u1 – d1)(u2 – d2) 121 u2 Ci > L /2 L u1* = u2* = L/2 C2 L/2 0 u* L/2 C1 L u1 122 u2 C2 < L /2 L u1* = L – C2 u2* = C2 L/2 C2 0 u* L/2 C1 L u1 123 Problem with the Nash solution: truncating the U set leaves the solution unchanged ! u2 u* 0 u1 124 Relative utilitarism (Kalai et Smorodinsky) Each individual has an aspiration level bi defined as the maximum utility level compatible with individual rationality. Nash solution does not depend on individual aspirations. Relative utilitarism consists in satisfying individual in proportion to their aspirations. This solution relies on an monotonocity axiom replacing Nash's axiom of independence with respect to irrelevant alternatives. 125 u2 b u* "idéal" point d2 d U d1 u1 126 u2 0 u1 127 u2 bi = Ci L ui* = Ci L C1 + C2 C2 u* 0 C1 L u1 128 2.4 Social welfare orderings 129 Welfarism postulates that the welfare of individuals is the only ingredient to be used to compare states of the world. Cardinal welfarism assumes that - individual welfare utilities are measured by a utility index - utilities can be "compared" Because it concentrates exclusively on utility profiles, welfarism has no ethical content. For instance, the "non-envy" criterion does not enter into account. The task of the benevolent dictator is to compare utility profiles (u1 ,..., un ) and to identify the best profile. 130 Efficiency-fitness is one of the basic concept of welfarism. It underlies utilitarism. Let A denote the set of feasible states. A state x in A Pareto-dominates a state x' in A if ui ( x) ui ( x ') for all i, with strict inequality for at least one i i.e. there is unanimity to move from state x to state x'. A feasible state x is Pareto optimal if it is not dominated by any feasible state. The other principle is compensation. It underlies egalitarism. 131 The preferences of the benevolent dictator are denoted by . It is called social welfare ordering and it is assumed to be complete and transitive. The most widely used are: - utilitarian: (u1,..., un ) - Nash: (u1,..., un ) (u1,..., un ) (u1,..., un ) u u i i u u i i - egalitarian (leximin): (u1 ,..., un ) where L (u1,..., un ) (v1,..., vn ) L (v1,..., vn ) is the lexicographic ordering applied to a reordering of the utility profiles in an increasing way. 132 Lifeboat Consider the case of 5 individuals and the following feasible arrangments: A {{1, 2},{1,3},{1, 4},{2,3,5}{3, 4,5},{2, 4,5}} Assume first that all individuals value equally being (10) and not being onboard (1). Utilitarism recommends choosing one of the 3-person arrangments. Egalitarism recommends the same solution: (1,1, 10,10,10) L (1,1, 1,10,10) 133 Assume now that utilities differ: in 1 10 2 6 3 6 4 5 5 3 out 0 1 1 1 0 Utilitarism now recommends choosing either {1,2} or {1,3}. The ranking is given by: {1, 2} {1,3} {1, 4} {2,3,5} {2, 4,5} {3, 4,5} 134 in 1 10 2 6 3 6 4 5 5 3 out 0 1 1 1 0 Egalitarism recommends {2,3,5}. Indeed the corresponding ranking is: {2,3,5} {2, 4,5} {3, 4,5} {1, 2} {1,3} {1, 4} obtained from: (0,1, 3, 6, 6) L (0,1, 3, 5, 6) L (0,1,1, 6,10) L (0,1,1, 5,10) 135 Collective utility function Most social welfare orderings can be represented by a collective utility function W(u1,…,un). A collective utility function W is additive if there exists some increasing function f such that: W (u1,..., un ) f (ui ) for all (u1,…,un). 136 Additive collective utility functions Social welfare orderings are assumed to be complete and transitive. Five additional assumptions 1. Monotonicity: ui ui for all i j and u j uj (u1,..., un ) (u1,..., un ) 2. Symmetry: if (u1,..., un ) is obtained from (u1 ,..., un ) by permuting individuals, then (u1 ,..., un ) (u1,..., un ) 137 Monotonicity is compatible with Pareto optimality: if (u1 ,..., un ) Pareto-dominates (u1,..., un ) then (u1 ,..., un ) (u1,..., un ) Hence, maximal elements on the set of feasible states A of a monotonic social welfare ordering are Pareto-optimal. Symmetry is equivalent to "equal treatment of equals": only differences in utilities may justify discrimination. 138 3. Ignoring unconcerned individuals: (ui , a) (u i , a) (ui , b) (u i , b) for all a, b where ui (u j | j i). Hence social welfare orderings depends only on the welfare of the individuals who are affected. Proposition Any social welfare ordering represented by an additive collective utility function satisfies the above property. Under continuity, the converse is true: ignoring unconcerned individual implies additivity. 139 4. Pigou-Dalton transfer principle: aversion for inequality If the utility profiles (u1 ,..., un ) and (u1,..., un ) are such that: u1 u2 ui ui for all i 1, 2 u1 u1 a and u2 u2 a then (u1,..., un ) (u1,..., un ). i.e. operating a transfer that reduces the inequality between any two individuals does not lead to a less preferred utility profile. 140 5. Independence of common scale A common rescaling of every individual utility function leaves the social welfare ordering unaffected: (u1,..., un ) (u1,..., un ) (u1,..., un ) (u1,..., un ) whenever 0 and uiui 0 for all i. Applied to an additive collective function, this property reads: f (u ) f (u) 0 i i f (u ) f (u) 0 i i Restricting to increasing and continuous functions f leads to… 141 Proposition Any additive, increasing and continuous social welfare ordering satisfying the invariance property (5) can be represented by a collective utility function of one of the following three types: W (u1 ,..., un ) uip for some p 0 f (u ) u p W (u1 ,..., un ) log ui f (u ) log u 1 W (u1 ,..., un ) p for some p 0 f (u ) u p ui 142 Maximizing log u is equivalent to maximizing u . Indeed, i i log is an increasing function and we have: log u i log ui Hence, log uiis called the Nash collective utility function. It is the limit of the other two families of utility function for p 0. The classical utilitarian utility function W (u1,..., un ) ui is obtained by setting p = 1 in the first family. 143 Proposition An additive utility function W (u1,..., un ) f (ui ) meets the Pigou-Dalton transfer principle if and only if the function f is concave. For instance, the quadratic utility function W (u1,..., un ) ui2 2 promotes inequality. Indeed, because ui ui transferring 2 utility to one individual is always preferable. 144 Conclusion: if we impose the five requirements - monotonicity and symmetry - ignoring unconcerned individuals - aversion for inequality - independance of common scale we are left with the following family of utility functions: W (u1 ,..., un ) uip for some p, 0 p 1 W (u1 ,..., un ) ui p for some p 0 including their limits for p 0. It is a one dimensional family defined by a single parameter p . 145 Leximin – egalitarian social welfare ordering Equalization of utilities may not be possible because the ranges of the utility functions differ. Equalization of utilities may be incompatible with Pareto efficiency. The leximin social welfare ordering selects the most egalitarian among the Pareto optimal allocations. 146 The leximin welfare ordering cannot be represented by a collective utility function. However, it belongs to the family of additive concave collective utility functions in a limit sense. Proposition The social welfare ordering represented by the collective utility function W (u1,..., un ) ui p converges to the leximin welfare ordering for p . 147 no equality – efficiency trade-off u2 u1 = u2 UT U(A) EG = LEX u1 + u2 = constant u1 148 no equality – efficiency trade-off u2 u1 = u2 u1 + u2 = constant U(A) EG = LEX UT u1 149 equality – efficiency trade-off u2 UT u1 = u2 LEX U(A) u1 + u2 = constant u1 150 u2 u1 = u2 NASH U(A) u1 u2 = constant u1 151 Independence of the common utility space The leximin ordering is invariant with respect to a common transformation of the utilities: (u1 ,..., un ) (T (u1 ),..., T (un )) L (v1 ,..., vn ) L (T (v1 ),..., T (vn )) Proposition Leximin is the only social welfare ordering satisfying the Pigou-Dalton transfer principle and the independence of the common utility space. 152 Independence of zero utilities The utilitarian social welfare ordering is invariant of zero utilities: (u1 ,..., un ) (v1 ,..., vn ) (u1 w1 ,..., un wn ) (v1 w1 ,..., vn wn ) for all (w1,..., wn ), or (u1 ,..., un ) (v1,..., vn ) (u1 v1,..., un vn ) (0,...,0) Proposition The utilitarian social welfare ordering is the only social welfare ordering satisfying independence of zero utilities. 153 Independence of utility scales The Nash social welfare ordering is independent of utility scales: (u1 ,..., un ) (v1 ,..., vn ) (a1u1 b1 ,..., anun bn ) (a1v1 b1 ,..., an vn bn ) for all ai 0 and bi . Proposition The Nash social welfare ordering is the only social welfare ordering satisfying independence of utility scales. 154 Example: location of a facility Consider the "linear" city represented by the interval [0,1] along which individuals are located: individual i is located at ti [0,1] If x denotes the location of the facility, the disutility of agent i is measured by its distance to the x: ui ( x) x ti 155 If there are agents located at 0 or 1, the egalitarian solution consists in placing the facility in the middle: x 1 / 2. The corresponding ordered utility vector is of the form (–1 /2, –1/2,….). It differs from the utilitarian solution which picks the median xˆ defined by: 1 {i | ti xˆ} and 2 1 {i | ti xˆ} 2 This is indeed the point where total disutility is minimum: moving away – in any direction – increases the disutility of at least 1/2 of the individuals. 156 Both solutions coincide when the individuals are uniformly distributed on the interval [0,1]. This is in particular the case of a continuum. The choice of the solution depends upon the kind of facility, in particular whether or not the facility is intented to meet basic needs (swimming pool vs post office). In some cases, the choice is difficult: where should a fire station be located ? 157 Example: location of a noxious facility Now, the distance to the facility measures the utility of agent i: ui ( x) x ti In the extreme case of a continuum, the egalitarian solution consists in locating the facility anywhere because there is an individual in any location. The utilitarian solution now picks one of the extreme points. The question is to compare the utilities at the end points. 158 Indeed, if f denotes the density function and is the mean, we have: 1 0 1 x f ( x)dx (1 x) f ( x)dx 2 1 0 Hence, location at 1 will be preferred if and only if 1 . 2 f(x) UT 0 1 x 159 Example: time sharing The problem is to share a given length of time between m radio programs to be broadcasted in a room where n individuals work. Each individual is assumed to either like or dislike a program: utilities are then either 0 or 1. Each program is supported by at least one individual. The problem is to allocate time in proportions t1,…,tm such that tk 0 for all k and t k 1 160 Assume first that each individual likes one and only one program and let nk denote the number of individuals who like program k: 0 nk n for all k and n k n Utilitarism implies majority: it picks the program supported by the largest group. In case where there is a tie, any combination is optimal. Egalitarism does the opposite: each program is broadcasted equally i.e. tk 1 for all k m 161 Assume now that individuals may be indifferent between radio programs. Consider the following case where n = m = 5: a b c d e 1 1 0 0 0 0 2 0 1 0 0 0 3 0 0 1 1 0 4 0 0 0 1 1 5 0 0 1 0 1 So as to equalize the portion of time each individual listen to a given program, egalitarism suggests the following allocation: 2 2 1 1 1 x( , , , , ) 7 7 7 7 7 162 a b c d e 1 1 0 0 0 0 2 0 1 0 0 0 3 0 0 1 1 0 4 0 0 0 1 1 5 0 0 1 0 1 Utilitarism instead suggest to forget about programs a and b, and to concentrate on programs c, d and e, with an arbitrary allocation. 163 If one particular program is supported by a majority, for instance: a b c d e 1 1 0 0 0 0 2 0 1 0 0 0 3 0 0 1 1 0 4 0 0 0 1 1 5 0 0 1 1 1 utilitarism would simply suggest to concentrate on that program, without paying attention to those outside that majority. 164 2.5 Transferable utility games 165 TU-games Given a collectivity N = {1,…,n}, a cooperative game with transferable utility is defined by a "characteristic function" v that associates a real number to any "coalition" S N. Here v(S) is the worth of coalition S, understood as the minimum it can secure for itself, independently of what the players outside S do. The set function v is assumed to be superadditive: S T v(S ) v(T ) v(S T ) a weaker requirement than convexity: S , T N v(S ) v(T ) v(S T ) v(S T ) 166 The problem is to share v(N) among the n players: find x = (x1,…,xn) such that x( N ) v( N ) The minimum requirements is individual rationality: xi v(i) for all i N This defines the set imputations: I ( N , v) {x n | x( N ) v( N ), xi v(i) for all i N} 167 The core extends the rationality requirement from individuals to coalitions: x(S ) v(S ) for all S N The core is the set, possibly empty, of allocations satisfying these conditions: ( N , v) {x n | x( N ) v( N ), x(S ) v(S ) for all S N} It is the set of allocations against which there can be no objections from any coalition, including individuals. Hence ( N , v) I ( N , v) 168 The core is not as such a solution. It is the set of "stable" allocations and there may be no such allocations except for some classes of games like for instance convex games. There are two "rules" that defines "fair" allocations. The Shapley value: it allocates v(N) on the basis of players marginal contributions to all coalitions they belong to: v( S ) v( S \ i ) It defines an imputation that may not belong to the core. The nucleolus: it selects an allocation that is always defined and belongs to the core when this one is nonempty. 169 Shapley value To each permutation = (i1,…,in) N of the players is associated a marginal contribution vector () defined by: i ( ) v(i1 ) v() v(i1 ) 1 i ( ) v(i1,..., ik ) v(i1,..., ik 1 ) k (k 2,..., n) The Shapley value is the average marginal contribution vector: 1 ( N , v) ( ) n! N 170 Alternatively, the Shapley value can be written as: i ( N , v) (s) [v(S ) v(S \ i)] SN ( S i ) n ( s 1)!(n s )! where the weights are given by n ( s ) n! The Shapley value is the unique allocation rule satisfying: - symmetry: contributions (substitute treatments of equals) players with identical marginal players) get the same ( equal - null player: players never contributing (null players) get nothing - additivity: (N,v+w) = (N,v) + (N,w) 171 1 1 n 2 2 ( , ) 2 2 2 (1,1) 1 1 1 n 3 3 ( , , ) 3 6 3 3 (1, 2,1) 1 1 1 1 n 4 4 ( , , , ) 4 12 12 4 4 (1,3,3,1) 1 1 1 1 1 n 5 5 ( , , , , ) 5 20 30 20 5 5 (1, 4,6, 4,1) 1 1 1 1 1 1 n 6 6 ( , , , , , ) 6 30 60 60 30 6 6 (1,5,10,10,5,1) where n ( s) Cns11 (n 1)! is the number of coalitions to which a given player belongs (n s)!( s 1)! n ( s) n ( s) 1 for all s n 172 Least core and nucleolus The Shapley value is "fair" because it treats equal players equally and does not remunerate non-contributing players. The nucleolus instead is concerned with reducing the highest loss of the coalitions as measured by the difference between wath a coalition is worth and what it gets: e( x, S ) v(S ) x(S ) is the "excess" associated to imputation x and coalition S. The least core is the set of imputations that minimize the largest excess: Min xI ( N , v ) Max S N e( x, S ) S , N 173 This is typically a set. The nucleolus goes further to eventually retain a unique imputation: to each imputation x is associated the vector (x) of dimension 2n – 2 obtained by placing the excesses e(x,S) in a decreasing order The nucleolus is then the unique imputations x that minimizes lexicographically these vectors on the set of imputations I(N,v): (x ) L ( x) for all x I ( N , v) 174 Example: "market" game v(1) = v(2) = v(3) = v(23) = 0 v(12) = p2 ≤ p3 v(13) = v(123) = p3 The core is defined by: ( N , v) { x ( p,0, p3 p) p2 p p3 } In particular, if p3 = p2, then ( N , v) { ( p3 ,0,0) } 175 1 2 3 123 0 200 100 132 0 0 300 213 200 0 100 231 300 0 0 312 300 0 0 321 300 0 0 1/6 1100 200 500 v(i) = 0 v(12) = p2 = 200 v(13) = p3 = 300 v(23) = 0 v(123) = p3 = 300 each row corresponds to a permutation chaque column corresponds to a player 550 100 250 ( N , v) ( , , ) (183,33,83) 3 3 3 176 For any given coalition, the excess can be written as a fucntion of p: e( p , S ) p 0 for S {1} for S {2} and S {13} p p3 for S {3} and S {23} p2 p for S {12} For each p, order the excesses in a decreasing way: (0, 0, p2 p, p p3 , p p3 , p) for p [ p2 , p] (0, 0, p p3 , p p3 , p2 p, p) for p [ p, p3 ] p2 p3 where p 2 177 p2 here the least core coincides with the core and the nucleolus is its mid-point p – p3 p p3 p 2 2 0 p2 p p3 (p2-p3)/2 p2-p3 p2 – p - p3 –p 178 The nucleolus is the mid-point of the core: p3 p2 p3 p2 ( N , v) ,0, 2 2 i.e (250, 0, 50) in the case where p3 = 300 and p2 = 200. The nucleolus satisfies to two Shapley's axioms: symmetry and nul player. It does not satisfy additivity. 179 Example: crop game Imagine a landlord and m (identical) workers, and a technology described by a production function y = F(s) where s is the number of workers: v(S) = 0 if S does not include the landlord v(S) = F(s – 1) if S includes the landlord (he/she does not work) In particular, v(i) = 0 for all i and v(N) = F(m). We suppose that F is increasing with F(0) = 0, not more at this stage. The associated game is superadditive. It is convex if returns to scale are constant or increasing: linear or convex production function. 180 We first observe that the extreme allocation (F(m), 0, …,0) always belongs to the core. Let x be in the core. For all j 1, we have: x( N \ j ) v( N \ j ) F (m 1) where x( N \ j) x( N ) x j F (m) x j Hence, x j F (m) F (m 1) the most a worker can get within the core is the marginal product [F(m) – F(m–1)] 181 Workers are substitutes: they get the same wage under the Shapley value. We need only to compute what the value allocates to the landlord. In a given permutation, only the position of the landlord counts. if the landlord is in position k, he gets F(k-1) and there are m + 1 positions possibles 1 ( N , v) m 1 m F (k 1) F (k ) m 1 m 1 1 k 1 1 k 1 182 F(m) F(1) + F(2) + … + F(m) F(k) 1 x F(k) F(2) F(1) 0 k k+1 m 183 decreasing returns F(m) Workers L>W Landlord m F (k ) k 0 0 m 184 constant returns F(m) Workers W=L Landlord 0 m 185 increasing returns F(m) T Workers L<W Landlord 0 m 186 mixed returns F(m) Workers Landlord 0 m 187 The Talmud example A man dies and his three wives have each a claim on his estate, following past promises. The value of the estate falls short of the total of the claims. Here is what a Mishnah suggests. d1=100 d2=200 d3=300 E=100 33.3 33.3 33.3 E=200 50 75 75 E=300 50 100 150 EQUAL ? UL 188 Aumann and Mashler (1985) have shown that the nucleolus actually reproduces the Talmud figures for the following TU-game: v(S ) Max 0, E d ( N \ S ) Here v(S) represents the minimum coalition S can get: it is the amount left once the outsiders have possibly got their claims In particular, v(N) = E. The above game is known as "bankcruptcy game". 189 E = 200 d = (100,200,300) v(S ) Max 0, E d ( N \ S ) v(i) = 0 i = 1,2,3 v(12) = v(13) = 0 v(23) = 100 v(123) = 200 x1 , x2 , x3 0 x1 x2 x3 200 x1 100 x2 x3 100 Here players 2 and 3 are substitutes. ( N , v) {( x1, x2 , x3 ) 0 x1 a, x2 x3 200 a, 0 a 100 } 190 x1 (200,0,0) 200 0 I ( N , v) x2 200 x3 200 x2 + x3 = 100 ( N , v) (0, 200,0) x2 + x3 = 200 x1 = 0 (0,0, 200) 191 E = 200 d = (100,200,300) v(1) = 0 v(2) = 0 v(3) = 0 v(12) = 0 v(13) = 0 v(23) = 100 v(123) = 200 v(S ) Max 0, E d ( N \ S ) 1 2 3 123 0 0 200 132 0 200 0 213 0 0 200 231 100 0 100 312 0 200 0 321 100 100 0 1/6 200 500 500 200 500 500 ( N , v) , , (33.7, 83.7, 83.7) 6 6 6 192 (200,0,0) (100, 100, 0) (100, 0, 100) Equal Nucleolus Shapley (0, 200,0) (0,0, 200) 193 We observe that the four vertices of the core are precisely the four marginal contribution vectors: (0, 0, 200) (0, 200, 0) (100, 0, 100) (100, 100, 0) with multiplicity 2 with multiplicity 2 with multiplicity 1 with multiplicity 1 This is actually a characteristic of convex games. Actually: the core of game is the convex hull of its marginal contribution vectors if and only if it is a convex game As a consequence, the Shapley value is in the core of convex games. The bankcruptcy game is convex. 194 E = 200 d1 = 100 d2 = 200 d3 = 300 EQUAL 66.6 66.6 66.6 PROP 33.3 66.6 100 UG 66.6 66.6 66.6 UL 0 50 150 Nucleolus 50 75 75 Shapley 33.3 83.3 83.3 195 E = 100 d = (100,200,300) v(S ) Max 0, E d ( N \ S ) v(i) = 0 i = 1,2,3 v(12) = v(13) = v(23) = 0 v(123) = 200 x1 , x2 , x3 0 x1 x2 x3 200 ( N , v) I ( N , v) The game is symmetric: all players are substitutes. i ( N , v ) i ( N , v ) 200 3 196 E = 300 d = (100,200,300) v(i) = 0 i = 1,2,3 v(12) = 0 v(13) = 100 v(23) = 200 v(123) = 300 v(S ) Max 0, E d ( N \ S ) x1 , x2 , x3 0 x1 x2 x3 300 x1 x3 100 x2 x3 200 197 (300,0,0) x1 + x3 = 100 x2 + x3 = 200 ( N , v) (0, 300,0) (0,0, 300) 198 E = 300 d = (100,200,300) v(i) = 0 i = 1,2,3 v(12) = 0 v(13) = 100 v(23) = 200 v(123) = 300 v(S ) Max 0, E d ( N \ S ) 1 2 3 123 0 0 300 132 0 200 100 213 0 0 300 231 100 0 200 312 100 200 0 321 100 200 0 1/6 300 600 900 300 600 900 ( N , v) , , (50, 100, 150) 6 6 6 199 (300,0,0) x1 + x3 = 200 x1 + x2 = 150 x1 + x3 = 100 x2 + x3 = 200 Equal Shapley = Nucleolus (0, 300,0) (0,0, 300) 200 (300,0,0) (100, 0, 200) (100, 200, 0) (0, 300,0) (0,200, 100) (0,0, 300) 201 1 2 3 123 0 0 300 132 0 200 100 213 0 0 300 231 100 0 200 312 100 200 0 321 100 200 0 We observe again that the four vertices of the core are precisely the four marginal contribution vectors: (0, 0, 300) (0, 200, 100) (100, 0, 200) (100, 200, 0) with multiplicity 2 with multiplicity 1 with multiplicity 1 with multiplicity 2 confirming that the bankcruptcy game is convex. 202 Assignment games (Shapley and Shubik) Consider a set N = {1,…,n} of agents and a set M = {1,…,m} (m n) of indivisible objects (say houses) to be allocated, one to each agent. Each agent attaches a "utility" to each house. These data are summarized in a utility matrix [ui (h) | i N , h M ] ui(h) is the reservation price of agent i for house h i.e. the maximum price i is willing to pay for house h. It is the value that agent i attach to house h expressed in monetary terms. 203 Side payments being allowed, the associated TU-game is given by: v(S ) Max f F iS ui ( f (i)) where F is the set of all functions f: N M that associates a house to each player. Here v(S) is the cost of the houses that are optimally allocated to the members of coalition S. Consequently, (N,v) is a cost game. It is concave and thereby also subadditive. 204 An optimal allocations of objects to players is associated to the definition of C(N) In the example below, it is (2,3,1): player 1 receives house 2, player 2 receives house 3, and player 3 receives house 1. C(1) = 12 1 u1 u2 u3 3 9 9 C(2) = 9 C(3) = 9 C(12) = 21 2 12 6 6 3 9 6 3 C(13) = 21 C(23) = 15 C(123) = 27 205 An allocation (y1,…yn) of C(N) specifies for each player the price he/she should pay for the object he/she has been assigned. The associated prices are (12,6,9) and the Shapley value of the game is given by (N,C) = (12, 7.5, 7.5) It implies the following side payments between players: (0, 1.5, – 1.5) i.e. player 1 stays put and player 2 pays 1.5 to player 3. 206 We observe that players 2 and 3 are substitute. The Shapley value is obtained from the following table which associates marginal cost vectors to players' permutations. 1 2 3 123 12 9 6 132 12 6 9 213 12 9 6 231 12 9 6 312 12 6 9 321 12 6 9 1/6 72 45 45 (N,C) = (12, 7.5, 7.5) 207 The core is defined by the allocations satisfying the following inequalities: y1 y2 y3 27 y1 12 y1 = 12 y2 9 6 y2 9 y3 9 6 y3 9 y1 y2 21 y1 y3 21 optimal allocation before transfers y2 y3 15 (12,6,9) (12,9,6) (12,7.5,7.5) the Shapley value is located at the center of the core 208 (12,9,6) (12,6,9) set of imputations (27,0,0) x2 = 6 x3 = 6 (9,9,9) x2 = 9 x3 = 9 x1 = 12 (0,27,0) (0,0,27) 209 3. Ordinal welfarism 210 A social choice procedure is a mapping F that associates alternatives to preference profiles: F : L( A)n A It associates to any profile p a subset of "winning" alternatives F(p) A. It is the collective choice set. A social welfare function is a mapping F that associates "collective" preferences to preference profiles: F : L( A)n L( A) 211 3.1 The case of two alternatives 212 Consider the case of 2 alternatives and n voters: A = {0,1} and N = {1,…,n} Assuming no indifference, a preference profile is a list of 0 and 1 of length n: p = (p1,…,pn) where pi L(A) = {0,1} where pi 1 1 i 0 pi 0 0 i 1 213 Example: n = 5 and p = (0,1,0,0,1) 3 in favour of 0 2 in favour of 1 There are 2n possible profiles. The set of all possible preference profiles is {0,1}n. A voting procedure is a mapping F: {0,1}n {0,1} It associates to any profile p a subset F(p) {0,1}. F(p) is the "choice set". 214 There are 4 possible outcomes: F(p) = {0} F(p) = {1} F(p) = {0,1} F(p) = So ties are allowed. The natural neutral mechanism to break a tie is the flipping of a coin. 215 Simple majority F ( p ) {1} F ( p ) {0} n n if pi 2 i 1 n n if pi 2 i 1 F ( p ) {0,1} if n pi i 1 n 2 a tie is not a possible outcome of simple majority if n is odd 216 Unanimity F ( p ) {1} n if p i 1 F ( p ) {0} if n p i 1 F ( p) i i n 0 n if 0 pi n i 1 217 A basic requirement to impose on a voting procedure is that it produces a result: Decisiveness A voting procedure is decisive if it never results in the empty outcome: F ( p) for all p {0,1}n Simple majority is always decisive. Unanimity is not. 218 What would be a fair voting procedure? What are desirable properties a voting procedure should have beyond decisiveness? The result of a voting procedure should not depend on the identity of the voters nor on the labelling of the alternatives: voters and alternatives should be treated equally 219 Anonymity A voting procedure F is anonymous if it symmetric in its n variables: for any p , permuting the voters leaves F(p) unchanged For instance, F (0,1,1,0,1) F (1,0,1,0,1) F (1,1,1,0,0) .... Anonymity clearly excludes dictatorship. It is actually a stronger form of non-dictatorship. 220 Neutrality A voting procedure F is neutral if permuting the choice of every voter results in a permutation of the outcome: for any p P, F(1 – p) = 1 – F(p) where 1 = (1,1,…,1). For instance, F (0,1,1,0,1) {1} F (1,0,0,1,0) {0} 221 Proposition: A voting procedure is anonymous and neutral if and only if it is the number of votes in favour of an alternative which determines whether he/she belongs to the choice set, i.e. n F ( p) G pi i 1 for some increasing function G. 222 Alone, anonymity and neutrality allow for many different voting procedures, including those based on stupid rules like: F ( p) {1} 1 1 n 4 if pi 10 n i 1 10 F ( p) {0} otherwise If an alternative is elected and some voters change their minds in favour of that candidate, it may be that he/she is not elected any more. 223 If, given the outcome F(p) corresponding to a preference profile p, some voters change their mind in favour of a candidate who belongs to the choice set F(p), we would expect that the resulting choice set still includes that alternative. Monotonicity A voting procedure is monotonic if 1 F ( p) and p p 1 F ( p) 0 F ( p) and p p 0 F ( p) where p p means pi pi for all i. 224 An increased support for an alternative should never hurt. An immediate consequence of monotonicity is strategyproofness: a voter has no incentive to be insincere by voting for the candidate he/she ranks second Is it possible to characterize the procedures which satisfy these 3 axioms simultaneously ? anonymity, neutrality and monotonicity 225 A quota procedure is defined by an integer F ( p) {1} n if p i 1 F ( p) {0} i q, n q n, such that: 2 q n if n pi q i 1 n p i 1 i nq F ( p) {0,1} otherwise Simple majority is defined by: n 1 q if n is odd 2 n q 1 if n is even 2 226 Unanimity is also a quota procedure with q = n. Proposition Quota procedures are the only voting procedures which are anonymous, neutral and monotonic. (i) quota procedures satisfy anonymity and neutrality by contruction: their outcome depends on the sum of the pi's. (ii) quota procedures satisfy monotonicity: F(p) does not decrease when the sum of the pi's increases. 227 A stronger version of the monotonicity axiom is the following: Strict monotonicity A voting procedure is strictly monotonic (positive responsiveness) if it monotonic and F ( p) {0,1} and p p F ( p) {1} F ( p) {0,1} and p p F ( p) {0} where p p means pi pi for all i and pj p j for some j. 228 If some voters change their mind in favour of a candidate who belongs to the initial choice set, then this alternative ends up being the only winner. In other words, either there was a tie and it disappears, or there was a unique winner and he/she remains the unique winner. Proposition Simple majority is the unique voting procedure which (May, 1952) is decisive, anonymous, neutral and strictly monotonic. 229 3.2 Social choice procedures 230 Borda method (1781) - each of the m position is graded: m – 1 for the 1st, m – 2 for the 2nd, …until 0 for the last - looking at the preference ordering of each voter, each alternative is graded accordingly - adding the grades, each alternative receives a score ... the alternative(s) with the largest score wins. 231 n=7 a b c d e Borda b a d b e c a d b e c a 14 c b d e a b 17 c d b a e c 16 b c d a e e c d b a d 16 e 7 232 Condorcet has criticized Borda's method. Consider 3 alternatives and 30 voters, 19 with preferences a b c 11 with preferences b c a For Condorcet, a should win while Borda assigns 41 to b against 38 to a. Indeed a is preferred to b and c by 19 voters. 233 An alternative is Condorcet winner if... ... confronted to any other alternative, it comes before in more than half of the orderings 1 a 2 a 3 a 4 c 5 c 6 b 7 e b b d b d c c c d b d b d d d e e e a a b e c c a e e a (This does not define a decisive rule !) 234 Hare method (1861) "single transferable voting system" - if an alternative comes on top of at least half of the orderings, he/she wins - if there is no such alternative, delete the alternative(s) that are on top of the fewest ordering - repeat the procedure with the remaining alternatives,... 235 a b c d e a d b e c a d b e c c b d e a c d b a e b c d a e e c d b a delete d 236 a b c e a b e c a b e c c b e a c b a e b c a e e c b a delete b and e 237 a c a c a c c a c a c a c a delete a Hare c 238 Sequential pairwise voting (voting with an agenda) The idea is that a sequence of alternatives is determined and followed. For instance, d results from the sequence (a,b,c,d,e) but b that comes out from the reverse sequence: a b c a d b a d b c b d c d b b c d e c d d e e c e c e a a e a e b a 239 Pareto criteria If all voters prefer x to y, then y cannot be in the social choice set. a a a c c b e b d d b a c c c b b a b a a d e e e d d b e c c d e e d 240 Condorcet criteria If there is a Condorcet winner, it must be in the social choice set. 1 a 2 a 3 a 4 c 5 c 6 b 7 e b b d b d c c c d b d b d d d e e e a a b e c c a e e a 241 Monotonicity criteria Let the alternative x be in the social choice set for a given preference profile p. If the preference profile p is modified by moving up x in the ordering of some voter,... ... x should remain in the social choice set. 242 Independence criteria (independence of irrelevant alternatives) Assume that the social choice set includes x but not y. If the preference profile P is modified, without altering the preferences between x and y,... ... then the resulting choice set should still not include y. 243 Pareto Condorcet Monotonicity Independance Plurality Yes No Yes No Borda Yes No Yes No Hare Yes No No No Agenda No Yes Yes No Dictator Yes No Yes Yes 244 Plurality satisfies Pareto If every voter prefers x to y, y cannot come on top of any ordering. Borda satisfies Pareto If x comes before y in all preference orderings, then x has more points than y. 245 Hare satisfies Pareto If every voter prefers x to y, y is not on top of any list. Then, either some alternative is on top of more than half of the orderings, it is the winner, not y, or y (being absent from the the first row) is among the alternatives to be deleted next. Dictatorship satisfies Pareto: if every voter prefers x to y, it is also the case of the dictator... 246 Sequential pairwise voting satisfies Condorcet If an alternative is the Condorcet winner, it will by definition come out of any sequence of pairwise votes. Plurality satisfies monotonicity If x be on top of the largest number of orderings, moving it up in some ordering preserves this. Borda satisfies monotonicity Moving up an alternative in some of the orderings always increases the number of points he/she gets... 247 Sequential pairwise voting satisfies monotonicity Assume x is a social choice given a preference profile and an agenda. Moving x up in the preferences of some voter will certainly keep x in the social choice set (with a larger margin). Dictatorship satisfies monotonicity If x is the social choice, it is on top of the dictator's ordering... Dictatorship satisfies independence If x is the social choice but not y, x is on top of the dictator's ordering and will remain so... 248 Plurality does not satisfy Condorcet 1 to 4 5 to 7 8 and 9 a b c b c b c a a a is plurality winner but b is Condorcet winner 249 Borda does not satisfy Condorcet 1, 2 and 3 a 4 and 5 b b c c a b is Borda winner but a is Condorcet winner 250 Hare does not satisfy Condorcet 1 to 5 a 6 to 9 e 10 to 12 d 13 to 15 c 16 and 17 b b b b b c c c c d d d d e e e e a a a a b is Condorcet winner but it will be deleted first 251 Dictatorship does not satisfy Condorcet 1 a 2 c 3 c b b b c a a c is Condorcet winner while a is the "social" choice if voter 1 is the dictator. 252 Hare does not satisfy monotonicity 1 to 7 a 8 to 12 c 13 to 16 b 17 b b a c a c b a c a is the social choice according to Hare 253 If voter 17 moves a above b, ... 1 to 7 a 8 to 12 c 13 to 16 b 17 a b a c b c b a c ... c becomes the social choice 254 Plurality does not satisfy independence a a b c b b c b c c a a a is the social choice and b is not 255 If voter 4 moves c between b and a, ... a a b b b b c c c c a a ... a and b are tied 256 Borda does not satisfy independence 1, 2 and 3 a 4 and 5 c b b c a a is the social choice 257 If voters 4 and 5 move c between b and a, ... 1, 2 and 3 a 4 and 5 b b c c a ... b becomes the social choice 258 Hare does not satisfy independence a a b c b b c b c c a a a is the social choice according to Hare 259 If voter 4 moves c between b and a, ... a a b b b b c c c c a a ... a and b are tied 260 Sequential pairwise voting does not satisfy Pareto a c b b a d d b c c d a b dominates d in the sense of Pareto: all voters prefer b to d but d results from the sequence (a,b,c,d) : a defeats b, c defeats a but d defeats c. 261 Sequential pairwise voting does not satisfy Independance c a b b c a a b c The reverse sequence (c,b,a) produces a as social choice. Interchanging c and b in the first ordering results in b as social choice while no one has changed his/her mind about a and b. 262 An illustration: Bonn, Berlin or both ? Bundestag, 20 June 1991 659 representatives, 3 alternatives: a = government in Bonn and parliament in Berlin b = government and parliament in Berlin c = government and parliament in Bonn A decision was eventually reached after a full day of debates. 263 Procedure adopted by the Council of Elders and the results: Abstention 18/654 147/654 Bonn and Berlin Yes End 489/654 No Abstention 29/657 Motion: NO to distinct locations 340/657 Yes No 288/654 338/659 Abstention 1/659 Berlin Bonn or Berlin 332/659 Bonn 264 Questions: Which voting procedure should have been adopted ? Does the actual voting procedure produce enough information to enable a reconstruction of the preferences of the 659 representatives ? Would a different voting procedure have produced a different outcome ? 265 Bonn-Berlin: Leininger's results* based on a clever reconstructed preference profile: 1. Majority would have been indecisive: 147/221/290. 2. Bonn would have been the plurality winner. 3. Berlin would have been the 2-step majority winner: 337/320. 4. Berlin is Condorcet winner: B/A: 371/286 B/C: 337/320 A/C: 227/430 5. Bonn would have been the Borda winner: B C A A = 513 B = 708 C = 750 6. Berlin and Bonn would have probably won under approval voting. *"The fatal vote: Bonn vs Berlin", Finanzarchiv, Neue Folge, Heft 1, 1993, 1-20 266 Scoring rules like Borda can be characterized. A scoring rule is defined by a mapping that associates weights to alternatives (assuming strict preferences) in terms of their positions in the preference lists. Consistency A social choice rule F is consistent if, for any two disjoint sets of voters N and N', and preference profiles p and p' on a common a set A of alternatives: F ( p) F ( p) F ( p) F ( p) F ( p p) where p p is the combined preference profile of N N'. 267 Proposition (Young) A voting procedure is anonymous, neutral and consistent if and only if it is a scoring rule. Remark: The Borda scoring rule has been axiomatized as well. 268 3.3 Impossibility theorems 269 Among the properties, the most desirable ones are certainly Pareto and monotonicity. Condorcet comes next. Independence appears as a strong requirement. It has indeed be the object of much discussion in the literature. We observe the following facts: - only dictatorship satisfies the independence axiom - only sequential pairwise voting satisfies the Condorcet axiom 270 Condorcet voting paradox a c b b a c c b a No Condorcet winner! Whatever is the social choice, 2/3 of the voters are unhappy and moreover, they agree on an other alternative ! 271 There is a transitivity problem! The collective preferences built by saying that "x is preferred to y" if and only if "x is preferred to y by a majority of voters" x y are not transitive: a b and b c but c a ... although individual preferences are. 272 One implicit assumption is made: there is no retrictions on the preferences: social choice function are defined for any preference profile in L(A)n The only requirement is that individual preferences are preorders. 273 Impossibility theorem 1 (Taylor) There is no decisive social choice procedure satisfying both the Condorcet and the independence criteria. Proof: - assume there exists such a procedure - apply it to the preference profile underlying the Condorcet paradox - show that it produces no winner: none of the three alternatives can be winning 274 Claim: a cannot be winning (same arguments for b and c) Consider the profile obtained from the Condorcet profile by moving b down in the third list: a c b b c a b c a a b c c a b c b a c is then Condorcet winner and must be in the choice set, not a. Going back to the Condorcet profile by moving b up in the third list should not affect the preferences between a and c. So a should still be a non-winner. 275 Impossibility theorem 2 (Arrow) Dictatorship is the only social welfare function satisfying the Pareto and independence criteria Impossibility theorem 3 (Gibbard) Dictatorship is the only social choice procedure satisfying the Pareto and monotonicity 276 These impossibility results remain true with a weaker Pareto requirement: If an alternative comes top in all preferences, then that alternative must be the unique social choice. This is indeed weaker than the original statement: if all voters prefer x to y, then y cannot be in the social choice set. 277 Strong monotonicity A social choice procedure F is strongly monotone if for all preference profile p and q in , and any alternative a in A: if q is obtained from p by lifting a up in some preference list, then either F(q) = F(p) or F(q) = a Pushing up an alternative can only help that alternative. Proposition (Muller and Satterthwaite) Dictatorship is the only social choice procedure satisfying the strong monotonicity 278 And what about the issue of manipulability: do voters have an incentive to report thruthfully their preferences i.e. to vote according to their preferences ? This is the problem known as strategic voting. It requires that voters have a fairly good idea of the preferences of the others. Non-manipulability could be one property that should be satisfied by a "good" social procedure. There too, there are impossibility results. 279 For a social procedure producing single outcomes, defining manipulability is easy. Let UA denote the set of utility functions representing complete preorders on A. Given a set A of alternatives, a social choice procedure F is nonmanipulable (or strategy-proof) if ui ( F (u)) ui ( F (vi , ui )) for all u U An , for all vi U A and for all i. 280 Example: Consider the Condorcet preference profile: a b c b c c a a b a b c b twice c twice a twice If a two step procedure is followed, a being opposed to b first we get: a vs b a a vs c c If voter 1 report the false preferences b be b instead. a c the outcome would 281 Impossibility theorem 4 (Gibbard and Satterthwaite) Assuming strict preferences, dictatorship is the only social choice procedure that is onto* and strategy-proof. The proof builds upon the impossibility theorem 3 according to which Pareto and monotonicity implies dictatorship: it is shown that a procedure satisfying strategy-proofness also satisfies Pareto and monotonicity. * i.e. surjective: each alternative can be winning for some preference profile. 282 Remark: In the case of 2 alternatives, strategy-proofness is equivalent to monotonicity and therefore, simple majority is strategy-proof. Other questions: Almost all social choice procedure are manipulable. The question could then be: is it possible to measure the degree of manipulability ? What about the possibility that a coalition of voters forms to jointly agree on a voting strategy ? 283 3.4 Possibility theorems 284 One implicit assumption is that social choice procedures are defined for any preference profile. We shall see that restricting the possible preference profiles results in possibility theorems. We shall consider two kinds of restrictions, Sen coherence and single peakedness, under which the collective preferences defined by x y if and only if a majority of voters prefer x to y are transitive. 285 Given a set A of m alternatives and a set N of voters, a preference profile p ( 1 ,..., i ,..., n ) is Sen coherent if, for all triplets (x,y,z) in A, one of the following three situations arises: x i y and x i z for all i N y i x and z i x for all i N y i x i z or z i x i y for all i N i.e. either x is preferred by all to y and z, or y and z are preferred to x by all, or all place x between y and z. 286 Given a set A of m alternatives and a set N of voters, a preference profile p ( 1 ,..., i ,..., n ) is single peaked if there exists an ordering of the alternatives such that each individual preference list has a peak. For 3 alternatives {a,b,c} and an ordering, say (b,a,c), individual preferences have a peak if one of the following situations arises: a b b a c c a b c a c b a b a b c c a b a c c b 287 Possibility theorem 1 (Sen) Collective preferences derived from Sen coherent preference profiles are transitive if n is odd. Possibility theorem 2 (Black) Collective preferences derived from single peaked preference profiles are transitive if n is odd. 288 Possibility theorem 3 (Moulin) If preferences are single-peaked and there is an odd number of voters, there is a (unique) Condorcet winner, the "median peak". Voters are ranked according to their peaks: a1 a2 ... an . The median peak is ak where k n 1 2 . There is a strict majority (k ) "leftists" (ai ak ). There is also a strict majority (k ) of "rightists" (ai ak ). Leftists support ak when opposed to a greater outcome. Rightists support ak when opposed to a smaller outcome. 289 Actually, the following proposition due to Moulin holds: Restricted to preference profiles involving an odd number of voters and for which a (unique) Condorcet winner (CW) exists, any social choice procedure producing the Condorcet winner is strategy-proof. It is also coalitionally strategy-proof: no coalition of voters can misreport its preferences and make its members better off. 290 Proof: Let D(A) be the set of individual preferences on A such that for all profiles p in D(A)n, CW(p) exists. Assume there exists a profile p in D(A)n, a coalition S in N and preferences qS in D(A)s such that: CW(p) = a and CW(qS,pN\S) = b a ui(a) < ui(b) for all i in S set of voters preferring a to b under profile p S N(a,b| p) = The set N(a,b| p) is a strict majority and the set N(a,b| (qS,pN\S)) coincides with the set N(a,b| p). Hence b cannot be Condorcet winner under the "false" profile (qS,pN\S). 291 Some interesting web sites: social choice: www.socialchoiceandbeyond.com game theory (non-cooperative and cooperative): www.citg.unige.it/siti_internet_web.html (in Italian) www.econ.canterbury.ac.nz/personal_pages/paul_walker (historical) arielrubinstein.tau.ac.il www.economics.utoronto.ca/osborne/igt (the site of his book) cooperative games: www.econ.usu.edu/acaplan/tugames.htm (a nice piece of software for n = 3) power indices: powerslave.val.utu.fi www.warwick.ac.uk/~ecaae (computation algorithms) 292