Abstract
I study the design of binary voting rules in environments where agents are heterogenous either in the stakes that they have in the decision or in the quality of information they possess regarding the correct course of action. The price of ‘one person, one vote’ is defined as the reduction in welfare resulting from constraining the voting rule to treat the agents symmetrically. I analyze how this price depends on the parameters of the environment, particularly on the level of heterogeneity in the society. The results shed light on the tradeoff between equality and efficiency in the context of constitutional design.
Similar content being viewed by others
Notes
According to Cicero, wealthier and older citizens had greater voting power to make the decisions of the assembly more in line with the will of those who are more experienced and had more to lose. See Abbott (1963, pp. 21, 75) for a description of the structure of the Centuriate Assembly before and after the reform.
While I term the constraint on the voting rule ‘one person, one vote’, the voters in the model can also be representatives of groups.
This document is available at http://www.fda.gov/downloads/RegulatoryInformation/Guidances/UCM125641.pdf.
Thus, in the private values model I assume that it is possible to measure intensity of preferences and make comparisons of this measure across agents. One possible interpretation of preference intensity in this context is in terms of willingness to pay: agents have quasi-linear preferences over pairs of an alternative and money, and the intensity of preferences is the maximal willingness to pay for their favorite alternative to be chosen. This is the view taken in the vast literature on costly voting, see Sect. 4 for references. In the context of international organizations, as described above, the intensity may reflect the size of a state or some measure of its relative importance in the world economy.
See Atkinson (1970) and Dasgupta et al. (1973) for a more modern treatment of inequality measures in general and the Pigou–Dalton principle in particular. This principle is at the core of inequality measurement: while many measures of inequality that completely rank distributions have been suggested in the literature, as far as I know they are all consistent with the Pigou–Dalton partial order.
To simplify notation I write a[n] and b[n] instead of a([n]) and b([n]), respectively.
The optimal welfare \(W^{*}\) is not everywhere differentiable and in the formal proof I use a different argument based on convexity. Nevertheless, the above argument provides the intuition behind the result.
The fact that \(\rho \) increases under these operations is an immediate consequence of Proposition 1 and the homogeneity of degree 0 of \(\rho \) as a function of a. However, since \(\tau \) is not homogeneous of degree 0 we cannot use Proposition 1 to conclude that it increases, and a different argument is needed to establish the result.
This type of worst-case analysis, which is most common in the Computer Science literature, has been used in several papers to measure the welfare loss due to constraints imposed on mechanisms, see for example Schmitz and Tröger (2006) in the context of voting, Neeman (2003) in auction design, and Pycia (2014) in matching markets. Note that for a worst-case analysis it is more informative to use the relative loss \(\rho \), since the absolute loss \(\tau \) is unbounded.
Note that the case \(0<\alpha _i<0.5\) is also covered by our analysis (by switching the labels of the signals h and l).
One way to see this formally is to fix some \(k<n/2\) and consider a bijection \(\varphi \) from the collection of coalitions of size k to the collection of coalitions of size \(n-k\), such that \(S\subseteq \varphi (S)\) for all S (such bijections can be explicitly constructed). Since \(\alpha _i>0.5\) for all i, it immediately follows that \(\alpha (S){\bar{\alpha }}(S^c) < \alpha (\varphi (S)) {\bar{\alpha }}(\varphi (S)^c)\). Summing up over all coalitions S of size k gives \(\sum _{\{S : |S|=k\}} \alpha (S){\bar{\alpha }}(S^c) < \sum _{\{S : |S|=k\}} {\bar{\alpha }}(S)\alpha (S^c)\), and hence that \(f(k)={\bar{L}}\) is optimal. For \(k>n/2\) a similar argument applies.
For this assertion to be true it is important that the expectation of \({\tilde{a}}_i\) is independent of the coalition S of agents who prefer alternative A. Correlation between the \({\tilde{a}}_i\)’s is allowed.
This can be interpreted as another reason for the use of anonymous mechanisms: if the voting rule cannot be changed frequently (i.e., the same rule should be used for many decisions), and if no agent systematically has higher stakes than others, then anonymous rules are optimal. I thank an anonymous referee for suggesting this argument.
See also the related result in Schmitz and Tröger (2006, Appendix B).
I thank an anonymous referee for suggesting this extension.
With different \(p_i\)’s the optimal anonymous rule is still qualified majority, but it is no longer necessarily true that when \(a=b\) simple majority is optimal. See Azrieli and Kim (2014, Theorem 4) for details.
References
Abbott FF (1963) A history and description of Roman political institutions. Biblo and Tannen, New York
Atkinson AB (1970) On the measurement of inequality. J Econ Theory 2:244–263
Azrieli Y, Jain R (2015) Symmetric mechanism design. Working paper
Azrieli Y, Kim S (2014) Pareto efficiency and weighted majority rules. Int Econ Rev 55:1065–1086
Bhattacharya S, Duffy J, Kim S (2015) Voting with endogenous information acquisition: theory and evidence. Working paper
Barberà S, Jackson MO (2006) On the weights of nations: assigning voting weights in a heterogeneous union. J Polit Econ 114:317–339
Ben-Yashar R, Milchtaich I (2007) First and second best voting rules in committees. Soc Choice Welf 29:453–486
Bognar K, Börgers T, Meyer-ter-Vehn M (2015) An optimal voting procedure when voting is costly. J Econ Theory 159:1056–1073
Börgers T (2004) Costly voting. Am Econ Rev 94:57–66
Dalton H (1920) The measurement of the inequality of incomes. Econ J 30:348–361
Dasgupta P, Sen A, Starrett D (1973) Notes on the measurement of inequality. J Econ Theory 6:180–187
Feddersen TJ, Pesendorfer W (1996) The swing voter’s curse. Am Econ Rev 86:408–424
Fleurbaey M (2008) Weighted majority and democratic theory. Working paper
Gerardi D, Yariv L (2008) Information acquisition in committees. Games Econ Behav 62:436–459
Gershkov A, Szentes B (2009) Optimal voting schemes with costly information acquisition. J Econ Theory 144:36–68
Goeree JK, Zhang J (2017) One man, one bid. Games Econ Behav 101:151–171
Groves T, Ledyard JO (1977) Optimal allocation of public goods: a solution to the “free rider” problem. Econometrica 45:783–809
Harstad B (2005) Majority rules and incentives. Q J Econ 120:1535–1568
Hilt E (2013) Shareholder voting rights in early American corporations. Bus Hist 55:617–632
Koriyama Y, Laslier JF, Macé A, Treibich R (2013) Optimal apportionment. J Polit Econ 121:584–608
Lalley SP, Weyl G (2015) Quadratic voting. Working paper
Marshal AW, Olkin I, Arnold BC (2011) Inequalities: theory of majorization and its applications, 2nd edn. Springer, New York
Martinelli C (2006) Would rational voters acquire costly information? J Econ Theory 129:225–251
May KO (1952) A set of independent necessary and sufficient conditions for simple majority decision. Econometrica 20:680–684
Neeman Z (2003) The effectiveness of English auctions. Games Econ Behav 43:214–238
Nitzan S, Parush J (1982) Optimal decision rules in uncertain dichotomous choice situations. Int Econ Rev 23:289–297
Okun AM (1975) Equality and efficiency: the big tradeoff. The Brookings Institution, Washington, DC
Oliveros S (2013) Abstention, ideology and information acquisition. J Econ Theory 148:871–902
Persico N (2003) Committee design with endogenous information. Rev Econ Stud 70:1–27
Pigou AC (1920) The economics of welfare. Macmillan, London
Pycia M (2014) The cost of ordinality. Working paper
Samuels SM (1965) On the number of successes in independent trials. Ann Math Stat 36:1272–1278
Schmitz PW, Tröger T (2006) Garbled elections. Working paper
Schmitz PW, Tröger T (2011) The (sub-)optimality of the majority rule. Games Econ Behav 74:651–665
Shapley LS, Grofman B (1984) Optimizing group judgmental accuracy in the presence of interdependencies. Public Choice 43:329–343
Author information
Authors and Affiliations
Corresponding author
Additional information
Y. Azrieli thank comments from participants of several seminars and conferences where this work has been presented. Special thanks to Marek Pycia, David Rahman, Maxim Ivanov, Eran Shmaya, and to three anonymous referees and an associate editor of Social Choice and Welfare for helpful suggestions.
Appendices
Appendix A: Proofs for Section 2
Proof of Lemma 3
From Lemma 1,
Now, for a fixed coalition S, the mapping \(a\longmapsto \max \left\{ a(S),a(S^c)\right\} \) is convex as the maximum of two linear functions. Since a convex combination of convex functions is also convex, it follows that \(W^*\) is a convex function of a. In addition, it is clear that \(W^*\) is symmetric in a, i.e., that a permutation of the coordinates of a does not change the value of \(W^*\). It is well-known [see, e.g., Marshal et al. (2011, Proposition C.2, p. 97)] that any function that is both convex and symmetric is monotonically non-decreasing with respect to the majorization partial order, which proves the assertion regarding \(W^*\).
Second, if a majorizes \(a'\) then \(a[n]=a'[n]\), so it follows immediately from Lemma 2 that \(W^{**}(a,p)=W^{**}(a',p)\). \(\square \)
Proof of Proposition 1
The proposition follows directly from Lemma 3. Indeed, if a majorizes \(a'\) then
and
Proof of Proposition 2
1. Given \(\delta >0\), let \(c^{\delta +} = \frac{a[n]}{a[n]+\delta } a^{\delta +}\). Recall that I assume in the proposition that a is ordered so that \(a_1\ge a_2\ge \cdots \ge a_n\), and notice that in moving from a to \(c^{\delta +}\) this order has not changed. For any \(k=1,\ldots ,n\),
with equality when \(k=n\). In other words, \(c^{\delta +}\) majorizes a, so it follows from Proposition 1 that \(\rho (c^{\delta +},p)\ge \rho (a,p)\). But recall that multiplying the utility functions of all agents by a constant does not affect \(\rho \), since both \(W^*\) and \(W^{**}\) are multiplied by that constant. Hence, \(\rho (c^{\delta +},p)= \rho (a^{\delta +},p)\). Combining these one gets
which is the required inequality for \(\rho \).
For \(\tau \), notice first that inequality (3) above can be rewritten as
Since clearly \(W^*(a^{\delta +},p)>W^*(a,p)>0\), I can rewrite this inequality as
where the last inequality follows from the definitions of \(W^*\) and \(W^{**}\). Rearranging gives
or
as needed.
2. The proof for \(\rho \) is similar to part 1 above. Define \(c^{\delta -} = \frac{a[n]}{a[n]-\delta } a^{\delta -}\). It is easy to check that \(c^{\delta -}\) majorizes a, and that \(\rho (c^{\delta -},p)= \rho (a^{\delta -},p)\). It then follows from Proposition 1 that
However, I cannot conclude in a similar way that \(\tau (a^{\delta -},p)\ge \tau (a,p)\), since here \(W^*(a^{\delta -},p)<W^*(a,p)\). Instead, I present a direct proof that does not rely on Proposition 1.
By Lemma 2,
Consider now the difference
The difference between the two maxima is positive only in the two cases where (1) \(n\in S\) and \(a(S)>a(S^c)\); or (2) \(n\in S^c\) and \(a(S^c)>a(S)\). In both cases the difference is at most \(\delta \). Therefore,
To bound this last expression, notice that since \(a_n\le a_i\) for any other agent \(i\in [n]\),
Summing up (5) over all \(i\in [n]\) and dividing by n gives
where the last equality is due to the fact that each winning coalition S of size k is counted k times in the sum. Repeating the argument, one obtains the inequality
Plugging (6) and (7) into (4) yields
Therefore, \(W^*(a,p) - W^*(a^{\delta -},p) \le W^{**}(a,p) - W^{**}(a^{\delta -},p)\), or \(\tau (a,p) \le \tau (a^{\delta -},p)\) as needed. \(\square \)
Proof of Proposition 3
From Lemma 2, if (a, b) sum-majorizes \((a',b')\) then \(W^{**}(a,b,p)=W^{**}(a',b',p)\). To prove the proposition I therefore need to show that \(W^*(a,b,p)\ge W^*(a',b',p)\). I do this by showing that a PD transfer from an agent with high stakes to an agent with lower stakes (as measured by the sum \(a_i+b_i\)) results in lower value of \(W^*\). As explained in the text, sum-majorization is characterized by finite sequences of such transfers, so this is sufficient to conclude that \(W^*(a,b,p)\ge W^*(a',b',p)\).
Fix the vectors a, b and suppose that \(a_i+b_i>a_j+b_j\). Let \(0<\delta < (a_i+b_i)-(a_j+b_j)\) and denote by \(a^\delta \) the vector obtained from a by replacing \(a_i\) by \(a_i-\delta \), replacing \(a_j\) by \(a_j+\delta \), and keeping all other coordinates unchanged. From Lemma 1,
For any coalition S that contains both i and j, or that contains neither of them, \(a^\delta (S)=a(S)\) holds. For coalitions S that contain i but not j one has \(a^\delta (S)=a(S)-\delta \). And for coalitions S that contain j but not i one has \(a^\delta (S)=a(S)-\delta \). Therefore, for each \(k=0,1,\ldots ,n\),
For every coalition S in the last sum (i.e., \(|S|=k\), \(j\in S,~ i\notin S\)), let \({\bar{S}}\) be defined by \({\bar{S}} = (S{\setminus } \{j\}) \cup \{i\}\). Note that \({\bar{S}}\) is in the first sum (i.e., \(|{\bar{S}}|=k\), \(i\in {\bar{S}},~ j\notin {\bar{S}}\)), and that the mapping \(S\longmapsto {\bar{S}}\) is one-to-one. Also, notice that \(a({\bar{S}}) = a(S)-a_j+a_i\) and \(b({\bar{S}}^c) = b(S^c)-b_i+b_j\), so
where the inequality is by the bound on \(\delta \).
I claim that for every such S
and I prove it by considering three possible cases:
Case 1: [\(a(S)\ge b(S^c)\)]. Here the right-hand side of (10) is equal to \(\delta \). From (9), \(a({\bar{S}})-b({\bar{S}}^c) \ge \delta \), which implies that the left-hand side of (10) is equal to \(\delta \) as well.
Case 2: [\(a(S) < b(S^c) \le a(S)+\delta \)]. In this case the right-hand side of (10) is equal to \(a(S)+\delta -b(S^c)\). As for the left-hand side, it follows again from (9) that \(a({\bar{S}})\ge b({\bar{S}}^c)\), so the left-hand side is equal to \(a({\bar{S}})-\max \{a({\bar{S}})-\delta ,b({\bar{S}}^c)\} = \min \{\delta , a({\bar{S}})-b({\bar{S}}^c)\} \ge \min \{\delta , a(S)-b(S^c)+\delta \} = a(S)-b(S^c)+\delta \). It follows that (10) holds in this case as well.
Case 3: [\(b(S^c)> a(S)+\delta \)]. In this last case the right-hand side of (10) is equal to 0. Since the left-hand side of (10) is clearly non-negative the proof is complete.
To conclude the proof of the proposition, note that (10) together with the fact that the mapping from S to \({\bar{S}}\) is one-to-one implies that (8) is non-negative for every k. This implies in turn that \(W^*(a,b,p)\ge W^*(a^\delta ,b,p)\) as needed. \(\square \)
Proof of Proposition 4
Fix a. To simplify notation I write \(g(p)=W^*(a,p)\) and \(h(p)=W^{**}(a,p)\), and view g, h as functions defined on the interval (0, 1) (suppressing the dependence on the fixed vector a). For every \(0\le k\le n\) let
Then it follows from Lemmas 1 and 2 that
Since \(g_k=g_{n-k}\) and \(h_k=h_{n-k}\) for all k, it follows that g and h are symmetric around \(p=0.5\), i.e. \(g(p)=g(1-p)\) and \(h(p)=h(1-p)\) for every \(p\in (0,1)\). To prove the proposition one therefore needs to show that both \(g-h\) and \(1-\frac{h}{g}\) increase for \(p\in (0,0.5)\).
Both g and h are polynomials and hence differentiable on (0, 1). In addition, g is bounded above zero on this interval, so the ratio h / g is differentiable. I show that \(h'(p)\le 0\) and \(h'(p)\le g'(p)\) for \(p\in (0,0.5)\), which is sufficient to prove the result. Indeed, from \(h'\le g'\) immediately follows that \(g-h\) is increasing. Also, since \(g\ge h>0\) it follows that \(h'g-g'h \le h'g-h'h = h'(g-h)\le 0\), which shows that \(\frac{h}{g}\) is decreasing.
To compute the derivatives \(g'\) and \(h'\), notice that for any function of the form \(f(p)=\sum _{k=0}^n p^k(1-p)^{n-k}f_k\) with \(f_k=f_{n-k}\) for all k the derivative is given by
where in the last equality \(f_k=f_{n-k}\) was used. Thus, to prove that \(h'\le 0\) and \(h' \le g'\) on (0, 0.5) it is sufficient to show that
and
for every \(k=0,\ldots ,\lfloor (n-2)/2\rfloor \). For every such k,
which proves (11).
For inequality (12), note that
For each coalition S (with \(|S|=k\)) and for each \(i\in S^c\) the above difference is bounded below by \(-a_i\). I therefore get that
which concludes the proof. \(\square \)
Proof of Proposition 5
First, since \(\rho \) is invariant to scalings of the utility functions, I may assume that \(a[n]=1\). Second, for any fixed p, \(W^*\) and \(W^{**}\) can be extended continuously from \(a\in \mathbb {R}^n_{++}\) to \(a\in \mathbb {R}^n_+\), and \(W^*(a,p)>0\) for any \(a\in \mathbb {R}^n_+\) except the origin. It follows that \(\rho \) can be extended continuously to the simplex \(\{a\in \mathbb {R}^n_+ :~ a[n]=1\}\). Now, Proposition 1 implies that for any given p the maximum of \(\rho \) with respect to a is attained at a maximal element of the majorization ordering, i.e. in one of the corners of the simplex. Proposition 4 then implies that \(\rho \) is maximal at \(p=0.5\). Plugging these values into the expressions for \(W^*\) and \(W^{**}\) yields
and
Simple manipulation gives
Thus, \(\bar{\rho } (n) = \frac{1}{2} - \left( \frac{1}{2}\right) ^n {n-1 \atopwithdelims ()\lceil \frac{n-1}{2}\rceil }\) as needed.
Finally, the expression \(\left( \frac{1}{2}\right) ^{n-1} {n-1 \atopwithdelims ()\lceil \frac{n-1}{2}\rceil }\) is the probability that out of \(n-1\) fair coin flips exactly \(\lceil \frac{n-1}{2}\rceil \) are heads. It is well-known and easy to see that this probability is decreasing in n and converges to 0. Therefore \(\bar{\rho } (n)\) is increasing in n and converges to \(\frac{1}{2}\). \(\square \)
Appendix B: Proofs for Section 3
For the proofs of this section it is convenient to introduce some auxiliary random variables and to express the relevant probabilities using these variables. Given a society \(\alpha \), let \(X_1,\ldots ,X_n\) be independent random variables with \(\mathbb {P}_\alpha (X_i=+1)=1-\mathbb {P}_\alpha (X_i=-1)=\alpha _i\). Let \({\bar{X}} = \sum _{i=1}^n X_i\). Also, if \(i\in [n]\) then I denote \({\bar{X}}_{-i} = \sum _{j\ne i} X_j\), and for \(i,j\in [n]\) I denote \({\bar{X}}_{-ij} = \sum _{l\ne i,j} X_l\).
Recall that \(\beta _i=\log \frac{\alpha _i}{1-\alpha _i}\). Let \(Y_1,\ldots ,Y_n\) be independent random variables with \(\mathbb {P}_\alpha (Y_i=\beta _i)=1-\mathbb {P}_\alpha (Y_i=-\beta _i)=\alpha _i\). Define \({\bar{Y}}\), \({\bar{Y}}_{-i}\) and \({\bar{Y}}_{-ij}\) in an analogous way.
Notice that the probabilities of a correct decision under the optimal and optimal anonymous rules are simply given by \(W^*(\alpha )=\mathbb {P}_\alpha ({\bar{Y}}> 0)+\frac{1}{2}\mathbb {P}_\alpha ({\bar{Y}}= 0)\) and \(W^{**}(\alpha )=\mathbb {P}_\alpha ({\bar{X}}> 0)\), respectively.
Proof of Lemma 6
I first prove monotonicity of \(W^{**}\). Notice that this function is continuously differentiable on \((0.5,1)^n\). Thus, by Marshal et al. (2011, Theorem A.4, p. 84), \(W^{**}\) is monotonic with respect to the majorization ordering if and only if it satisfies the condition
at any point \(\alpha \) and for any pair of agents i, j. I compute the partial derivatives of \(W^{**}\) and show that (13) holds.
Given \(\alpha \) and i,
Hence,
where the last equality follows from the assumption that n is odd. In words, the marginal increase in the probability of a correct decision with respect to \(\alpha _i\) is the probability that i is pivotal, i.e., the probability that exactly half of the other agents get an h signal. Now, given two agents i and j, the difference between the corresponding partial derivatives is
Thus, to prove that (13) holds it is enough to show that \(\mathbb {P}_\alpha ({\bar{X}}_{-ij} =1)\ge \mathbb {P}_\alpha ({\bar{X}}_{-ij} =-1)\) for any pair of agents i and j. By Samuels (1965, pp. 1272–1274), the distribution of \({\bar{X}}_{-ij}\) is unimodal, first increasing, then decreasing, and the mode can be either a single odd integer or a pair of adjacent odd integers. Furthermore, since \(\alpha _l>0.5\) for all agents, it follows from Theorem 2 in Samuels (1965) that any mode must be positive. Hence, \(\mathbb {P}_\alpha ({\bar{X}}_{-ij} =1)\ge \mathbb {P}_\alpha ({\bar{X}}_{-ij} =-1)\) which concludes the proof for \(W^{**}\).
Moving on to \(W^*\), I cannot immediately use condition (13) since \(W^*\) is not everywhere differentiable. Specifically, \(W^*\) is continuously differentiable only on the set \(D=\{\alpha ~:~ \mathbb {P}_\alpha ({\bar{Y}}=0)=0\}\), as can be seen from Lemma 4. I claim however that if \((\alpha _i -\alpha _j) \left( \frac{\partial W^*}{\partial \alpha _i}-\frac{\partial W^*}{\partial \alpha _j}\right) \ge 0\) everywhere in D then the proposition will be proved. Indeed, given two agents i and j, let \(d\in \mathbb {R}^n\) be defined by \(d_i=1\), \(d_j=-1\) and \(d_l=0\) for any other agent l. By Marshal et al. (2011, Lemma A.2, p. 81), if \(\alpha _i\ge \alpha _j\) implies that the mapping \(t\longmapsto W^*(\alpha +t d)\) is non-decreasing for \(t\ge 0\) (where it’s defined), then the proof is complete. But it is immediate to check that for every \(\alpha \) the interval \(\{\alpha +t d : t\ge 0\}\) intersects \(D^c\) only finitely many times. Thus, the function \(t\longmapsto W^*(\alpha +t d)\) is continuously differentiable at all but finitely many points on its domain (and continuous everywhere), so to show that it is non-decreasing it is enough to show that its derivative is non-negative wherever the derivative exists. But this is equivalent to showing that \(\frac{\partial W^*}{\partial \alpha _i} \ge \frac{\partial W^*}{\partial \alpha _j}\) in D.
Now, fix \(\alpha \in D\) and notice that
Since \(\alpha \in D\), \(\mathbb {P}_\alpha (-\beta _i-\epsilon< {\bar{Y}}_{-i}< -\beta _i+\epsilon ) = \mathbb {P}_\alpha (\beta _i-\epsilon< {\bar{Y}}_{-i} < \beta _i+\epsilon )=0\) for small enough \(\epsilon >0\), and hence \(\frac{\partial }{\partial \alpha _i}\mathbb {P}_\alpha ({\bar{Y}}_{-i}> -\beta _i) = \frac{\partial }{\partial \alpha _i}\mathbb {P}_\alpha ({\bar{Y}}_{-i} > \beta _i)=0\). Thus,
which, as in the anonymous case above, is the probability that agent i is pivotal.
Notice that \(\alpha _i\ge \alpha _j\) implies that \(\beta _i\ge \beta _j\), and therefore that \(-\beta _i+\beta _j\le 0 \le \beta _i-\beta _j\). Hence, \(\alpha _i\ge \alpha _j\) implies that
so it is sufficient to show that
Notice that \({\bar{Y}}_{-ij}=\beta (S)-\beta (S^c)\), where \(S=\{l\ne i,j : Y_l=\beta _l\}\) and the complement is relative to the set \([n]{\setminus } \{i,j\}\). Also, recall that \(\beta (S)<\beta (S^c)\) is equivalent to \(\alpha (S){\bar{\alpha }}(S^c) < {\bar{\alpha }}(S)\alpha (S^c)\). Therefore, any realized S at which \({\bar{Y}}_{-ij}=\beta (S) -\beta (S^c)<0\) has smaller probability than the realization \(S^c\) (at which \({\bar{Y}}_{-ij} =\beta (S^c)-\beta (S)>0\)). It follows that for every \(0<c<d\), \(\mathbb {P}_\alpha (c< {\bar{Y}}_{-ij}< d ) \ge \mathbb {P}_\alpha (-d< {\bar{Y}}_{-ij} < -c )\), which proves (14) \(\square \)
Proof of Proposition 6
1. Fix \(\alpha _1\ge \cdots \ge \alpha _{n-1}\). I prove that \(\tau (\alpha ) = W^*(\alpha ) - W^{**}(\alpha )\) is (weakly) decreasing in \(\alpha _n\in (0.5, \alpha _{n-1})\). Note that \(\tau \) is a continuous and piece-wise linear function of \(\alpha _n\), and that it’s derivative exists at any point in which \(\mathbb {P}_\alpha ({\bar{Y}}=0)=0\). Thus, it is sufficient to show that \(\frac{\partial W^*}{\partial \alpha _n} \le \frac{\partial W^{**}}{\partial \alpha _n}\) at any such point. From the proof of the previous lemma one has that \(\frac{\partial W^*}{\partial \alpha _n} = \mathbb {P}_\alpha (|{\bar{Y}}_{-n}| < \beta _n)\) and that \(\frac{\partial W^{**}}{\partial \alpha _n} = \mathbb {P}_\alpha ({\bar{X}}_{-n} =0)\). Thus, one needs to prove that \(\mathbb {P}_\alpha (|{\bar{Y}}_{-n}| < \beta _n) \le \mathbb {P}_\alpha ({\bar{X}}_{-n} =0)\), i.e., that the probability that agent n is pivotal under simple majority is at least as large as the probability that he is pivotal under the optimal rule.
The argument is based on the following two claims, whose proofs follow the proof of the proposition.
Claim 1
Let M be a set with an even number m of elements and let \(\Gamma \) be a collection of subsets of M such that no two different sets in \(\Gamma \) contain each other. Then there exists a one-to-one mapping f from \(\Gamma \) to \(\Gamma ':= \{T\subseteq M ~:~ |T|=m/2\}\) such that \(S\subseteq f(S)\) or \(f(S)\subseteq S\) for every \(S\in \Gamma \).
Claim 2
Let \(S\subseteq [n]{\setminus }\{n\}\) satisfy \(|\beta (S)-\beta (S^c)|<\beta _n\) (complements are relative to the set \([n]{\setminus }\{n\}\)). Then for every coalition \(T\subseteq [n]{\setminus }\{n\}\) of size \(|T|=\frac{n-1}{2}\) satisfying \(T\supseteq S\) or \(T\subseteq S\), \(\alpha (S){\bar{\alpha }}(S^c)+\alpha (S^c){\bar{\alpha }}(S) \le \alpha (T){\bar{\alpha }}(T^c)+\alpha (T^c){\bar{\alpha }}(T)\).
Let \(\Gamma =\{S \subseteq [n]{\setminus }\{n\} ~:~ |\beta (S)-\beta (S^c)|< \beta _n\}\) and \(\Gamma '=\{T \subseteq [n]{\setminus }\{n\} ~:~ |T| = \frac{n-1}{2}\}\). Note that if \(S\in \Gamma \) and \(S\subsetneqq {\tilde{S}}\) then
where the first inequality follows from \(S\in \Gamma \) and the last inequality from \(\beta _n \le \beta _i\) for all \(i\in [n]{\setminus }\{n\}\). It follows that \(\Gamma \) satisfies the condition of Claim 1, so there is a one-to-one mapping \(f:\Gamma \rightarrow \Gamma '\) such that \(S\subseteq f(S)\) or \(f(S)\subseteq S\) for every \(S\in \Gamma \). Therefore,
where the first equality follows from the definition of \(\Gamma \), the second equality from the fact that \(\Gamma \) is closed under complements, the first inequality from Claim 2 and the fact that either \(S\subseteq f(S)\) or \(f(S)\subseteq S\) for every \(S\in \Gamma \), the second inequality from the fact that f is one-to-one, the next equality from the fact that \(\Gamma '\) is closed under complements, and the final equality from the definition of \(\Gamma '\). This completes the proof that \(\tau \) is non-increasing in \(\alpha _n\) over \((0.5, \alpha _{n-1})\). Since \(\rho (\alpha ) = \frac{\tau (\alpha )}{W^*(\alpha )}\), and since \(W^*\) is weakly increasing in each of its arguments, it follows that \(\rho \) is also non-increasing in \(\alpha _n\) over \((0.5, \alpha _{n-1})\).
2. For an example in which \(\rho (\alpha ^{\delta +})>\rho (\alpha )\) and \(\tau (\alpha ^{\delta +})>\tau (\alpha )\) simply take \(\alpha \) to be a constant vector. Then \(\rho (\alpha )=\tau (\alpha )=0\), but after increasing \(\alpha _1\) sufficiently both measures become positive.
I now give an example in which \(\rho (\alpha ^{\delta +})<\rho (\alpha )\) and \(\tau (\alpha ^{\delta +}) <\tau (\alpha )\). Let \(n=7\) with \(\alpha _1=\alpha _2=\alpha _3=\alpha _4=1-\epsilon \) and \(\alpha _5=\alpha _6=\alpha _7=0.5+\epsilon \) for small \(\epsilon >0\). I show that slightly increasing \(\alpha _1\) decreases both \(\rho \) and \(\tau \).
Consider first the partial derivative of \(W^*\) with respect to \(\alpha _1\) (which exists since \(\mathbb {P}_\alpha ({\bar{Y}}=0)=0\)). As has already been shown, this derivative is equal to the probability \(\mathbb {P}_\alpha (|{\bar{Y}}_{-1}| < \beta _1)\). But notice that if agents 2, 3 and 4 all got an h signal (i.e., \(Y_2 =Y_3=Y_4>0\)) then \({\bar{Y}}_{-1} > \beta _1\) regardless of the signals of the other agents, so agent 1 is not pivotal. Thus,
The partial derivative of \(W^{**}\) with respect to \(\alpha _1\) is equal to \(\mathbb {P}_\alpha ({\bar{X}}_{-1}=0)\). One possible realization yielding \({\bar{X}}_{-1}=0\) is \(X_2=X_3=X_4=1\) and \(X_5=X_6=X_7=-1\). Therefore,
It follows that \(\frac{\partial W^*}{\partial \alpha _1} < \frac{\partial W^{**}}{\partial \alpha _1}\) whenever \(\epsilon >0\) is sufficiently small. This implies that \(\tau (\alpha ^{\delta +}) < \tau (\alpha )\) for \(\delta >0\) sufficiently small. In addition, for every such \(\delta \), \(\rho (\alpha ^{\delta +}) = \frac{\tau (\alpha ^{\delta +})}{W^*(\alpha ^{\delta +})} < \frac{\tau (\alpha )}{W^*(\alpha )} = \rho (\alpha )\). \(\square \)
Proof of Claim 1
The proof of the claim is based on the following fact. For any \(1\le k< m/2\) denote \(\Lambda _k=\{S\subseteq M ~:~ |S|=k\}\). Then for every such k there exists a one-to-one mapping \(g_k:\Lambda _k \rightarrow \Lambda _{k+1}\) such that \(S\subseteq g_k(S)\) for all \(S\in \Lambda _k\). The existence of such mappings can be proved by induction on m.
Now, consider a collection \(\Gamma \) which satisfies the condition in the claim. As a first step I argue that I may assume that \(\Gamma \) only contains sets of cardinality less than m / 2. Indeed, suppose that the claimis proved under this additional requirement. Then given a general collection \(\Gamma \) one can partition \(\Gamma \) into the three sets \(\Gamma _1=\{S\in \Gamma ~:~ |S|<m/2\}\), \(\Gamma _2=\{S\in \Gamma ~:~ |S|=m/2\}\) and \(\Gamma _3=\{S\in \Gamma ~:~ |S|>m/2\}\). Let \(f_1\) be the required function for \(\Gamma _1\). Let \(f_2\) be defined by \(f(S)=S\) for any \(S\in \Gamma _2\). Let \(f_3\) be the function corresponding to \(\{S^c ~:~ S\in \Gamma _3\}\). Then it is easy to check that the function f defined by \(f(S)=f_1(S)\) for \(S\in \Gamma _1\), \(f(S)=f_2(S)\) for \(S\in \Gamma _2\), and \(f(S)=f_3(S^c)\) for \(S\in \Gamma _3\) satisfies the requirement for the entire collection \(\Gamma \).
Thus, suppose that \(\Gamma \) contains only sets of cardinality less than m / 2 and satisfies the condition in the claim. For each \(S\in \Gamma \) define \(f(S) = g_{\frac{m}{2}-1}( \ldots ( g_{|S|+1} ( g_{|S|} (S))))\). Then f maps \(\Gamma \) into \(\Gamma '\) and clearly \(S\subseteq f(S)\) for all \(S\in \Gamma \). Also, it follows from the properties of \(\{g_k\}_k\) and from the assumption that no two sets in \(\Gamma \) contain one another that f is one-to-one. This concludes the proof. \(\square \)
Proof of Claim 2
Let S and T be as in the claim and suppose that \(S\subseteq T\) (the case \(T\subseteq S\) is analogous). If \(S=T\) then the claim is trivially true, so suppose that S is a strict subset of T. Then simple algebra gives
Since \(\alpha _i>1-\alpha _i\) for every i it follows that \({\bar{\alpha }} (T{\setminus } S) < \alpha (T{\setminus } S)\), so to complete the proof I need to show that \(\alpha (S) {\bar{\alpha }}(T^c) \ge {\bar{\alpha }}(S) \alpha (T^c)\). Rearranging and taking \(\log \) of both sides shows that this is equivalent to \(\beta (S) \ge \beta (T^c)\). This latter inequality follows from
where the first inequality follows from the assumption of the claim and the last inequality from the assumption that S is a strict subset of T and that \(\beta _n\le \beta _i\) for every agent i. \(\square \)
Proof of Proposition 7
The functions \(W^*(\alpha )\) and \(W^{**}(\alpha )\) are continuous on \((0.5,1)^n\) and can be continuously extended to the compact set \([0.5,1]^n\). I prove that the maximum of \(\tau \) on this set is attained at \((1,0.5,\ldots ,0.5)\).
Fix an agent i and \(\alpha _{-i}\). Then \(W^{**}(\cdot ;\alpha _{-i})\) is a linear function of \(\alpha _i\) with slope \(\mathbb {P}_\alpha ({\bar{X}}_{-i} =0)\), and \(W^*(\cdot ;\alpha _{-i})\) is piece-wise linear in \(\alpha _i\) with slope \(\mathbb {P}_\alpha (|{\bar{Y}}_{-i}| < \beta _i)\) at any point \(\alpha _i\) in which it is differentiable (see the proof of Lemma 6). Since \(\beta _i\) is monotonically increasing in \(\alpha _i\) it follows that \(W^*(\cdot ;\alpha _{-i})\) is convex. Therefore, the difference \(\tau (\cdot ;\alpha _{-i}) = W^*(\cdot ;\alpha _{-i}) - W^{**}(\cdot ;\alpha _{-i})\) is a convex function of \(\alpha _i\). This implies that the maximum of this function is attained at one of the boundaries, either at \(\alpha _i=0.5\) or at \(\alpha _i=1\).
It follows that \(\tau (\alpha )\) is maximized at a point in which each agent is either perfectly informed or completely uninformed. At \(\alpha =(0.5,\ldots ,0.5)\) one has \(\tau (\alpha )=0\), so this cannot be a maximizer. At any other corner of the cube at least one agent is perfectly informed, so that \(W^*(\alpha )=1\). Since \(W^{**}\) is increasing in each of its arguments, it follows that it is minimal when only one agent, say agent 1, is perfectly informed and everyone else is uninformed. At this point \(\alpha \),
Thus, \({\bar{\tau }} (n) = \frac{1}{2} - \left( \frac{1}{2}\right) ^n {n-1 \atopwithdelims ()\frac{n-1}{2}}\) as needed. The rest of the proof is the same as in the proof of Proposition 5. \(\square \)
Appendix C: Heterogeneous probabilities example
Consider a society of \(n=4\) agents with \(a_1=4\), \(a_2=a_3=3\), \(a_4=1\), and \(b_i=a_i\) for all i. Let the vector of probabilities be \((p_1,p_2,p_3,p_4)=(1-\epsilon ,0.5,0.5,\epsilon )\), where \(\epsilon >0\) is small. I compare \(\tau \) and \(\rho \) between this society and the society with \(a^{\delta +}\), i.e., the society in which \(a_1=4+\delta \) and everything else is unchanged.
Note first that whenever at least three of the agents support the same alternative (either A or B) then it is optimal to choose that alternative. It follows that the optimal rule and the optimal anonymous rule coincide at any such realization of types. So the rules may differ only if two agents support A and the other two support B. Now, since agent 1 is much more likely to prefer A and agent 4 is much more likely to prefer B, it follows that conditional on a tie (two agents support each alternative) the optimal anonymous rule chooses A. On the other hand, the optimal rule chooses A when the pair that prefers A is either \(\{1,2\}\) or \(\{1,3\}\) or \(\{2,3\}\), and chooses B otherwise. Therefore,
When moving to \(a^{\delta +}\), as long as \(\delta <1\) the optimal and optimal anonymous rules do not change. A similar computation then gives
This also implies that
Hence, increasing the inequality of stakes in the society resulted in a decrease of both measures of the price of OPOV, contrary to the conclusions of Sect. 2.3. \(\square \)
Rights and permissions
About this article
Cite this article
Azrieli, Y. The price of ‘one person, one vote’. Soc Choice Welf 50, 353–385 (2018). https://doi.org/10.1007/s00355-017-1087-z
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00355-017-1087-z