TSE‐511

July 2014

“Influence Vs. Utility in the Evaluation of Voting Rules: A

New Look at the Penrose Formula”

Michel Le Breton and Karine Van der Straeten

Influence Vs. Utility in the Evaluation of Voting Rules:
A New Look at the Penrose Formula
Michel Le Breton∗ Karine Van der Straeten†
July 2014

Abstract
In this paper, we clarify the relationship between influence/power measurement and
utility measurement, the most popular two social objective criteria used when evalu-
ating voting mechanisms. For one particular probabilistic model describing the pref-
erences of the electorate, the so-called Impartial Culture (IC) model used by Banzhaf,
the Penrose formula show that the two objectives coincide. The IC probabilistic model
assumes that voter preferences are independent. In this note, we prove a general ver-
sion of the Penrose formula, allowing for correlations in the electorate, and show that
in that case, the two social objectives no longer coincide and qualitative conclusions
can be very different.

1 Introduction
The purpose of this note is to clarify the relationship between power measurement and utility
measurement/voting design in the case where a group (society, assembly, committee,...) must
decide among two alternatives. Power measurement is a developed and popular area in
applied political science which has already a long history and has received a great deal of
attention while voting design (which we view as the application of mechanism design to
the normative analysis of political institutions) is newer. One possible explanation of this
discrepancy is the implicit belief that if a player is influential (powerful, pivotal,...) in an
institution then her utility will also be large as she will be in position to reduce the gap
between the collective decision and her favorite one. While intuitive, this assertion is not
true in full generality and the relationship between the two notions will depend upon the
probabilistic model describing the preferences of the electorate.
∗
Institut Universitaire de France and Toulouse School of Economics, France.
†
Toulouse School of Economics, CNRS and Institute for Advanced Study in Toulouse (IAST), France.

1
 The most popular two answers to the first question (what is the probability that my vote
will make a difference in the group decision under a given mechanism?) are the power indices
due to Banzhaf (1965) and Shapley and Shubik (1954); they are defined as the probability
for a player to be influential for two different probabilistic models. Sensitivity is defined1 by
Felsenthal and Machover (1998) as the sum of the power indices over the individuals. To
defend this quantity as a reasonable social objective, they write ”the sum of power indices
can be regarded as a measure of the sensitivity of the decision rule : the ease with which it
responds to fluctuations in the voters’ wishes”.
In contrast to the first question, the second question (what is my expected utility under
a given mechanism?) has been somehow neglected and the unique attempt to measure the
satisfaction or utility of a player in a voting body is due to Rae (1969) who has proposed
an index of satisfaction. Likely, one reason for this neglect is the nice relationship between
satisfaction and power that was established earlier2 by Penrose (1946) for the particular
probabilistic model used by Banzhaf, the so-called Impartial Culture (IC) model. He shows
that in such a case the satisfaction index is an affine transform of the power index3 . Both
Rae and Penrose suppose that the utility of any of the two alternatives can only take two
values, which are the same for all the individuals. Note that in such a case, the utility of
any player coincides with the probability that the group decision will agree with the player
decision.4
In this note, we prove a general (while still assuming that the utility takes only two
values) version of the Penrose formula which highlights the differences between power and
utility measurement in a general probabilistic model. It calls the attention on the role of
correlation as the main explanation of the gap between the two notions. After deriving
a general formula for any random electorate (section 3), we offer a detailed calculation
of total expected utility attached to any mechanism in the probabilistic model underlying
the Shapley-Shubik power index, the so-called Impartial Anonymous Culture (IAC) model
1
Definition 3.3.1. on page 52.
2
Felsenthal and Machover (1998) report on the history of the result. The formula is also proved in Brams
and Lake (1978) and Dubey and Shapley (1979). Dubey and Shapley comment that the connection between
satisfaction and the Banzhaf’s index was not noticed for several years after 1969. In fact, the connection has
been noticed a good many years before, in 1946; it is stated in Penrose without proof, something the readers
are expected to work out for themselves.
3
Penrose’s formula is quoted as theorem 3.2.16 in Felsenthal and Machover (1998). They reproduce the
following nice observation by Penrose: ” In general, the power of the individual vote can be measured by
the amount by which his chance of being on the winning side exceeds one half. The power, thus defined, is
the same as half the likelihood of a situation in which an individual can be decisive...”.
4
The extension of the Banzhaf’s probabilistic model to more than two alternatives (known also as the IC
model) has been used by several authors including for instance Myerson (1998, 2000) and Weber (1995).

2
(section 4). We conclude that while the results obtained through IC are insightful, we should
be careful about evaluating the qualities of alternative social mechanisms and deriving the
optimal one on the sole basis of the IC model, since qualitative conclusions can be very
different when allowing for some correlations in the electorate.

2 Random Electorates and Voting Mechanisms

We consider a society N = {1, 2, ..., n} of voters facing a choice between two alternatives
A and B. The state space is the set of profiles of strict preferences. A profile of strict
preferences is a vector X ∈ {0, 1}N where Xi = 1 (respectively 0) means that voter i prefers
A to B (prefers B to A). We will assume hereafter that the utility can only take two values:
1 for the best alternative and 0 for the worst.
A random electorate is a joint probability distribution λ over {0, 1}N . Hereafter, we will
denote Pi = Prλ [Xi = 1] the probability that voter i prefers A to B.
A random electorate λ is uniform if Pi = Pj = P for all i, j ∈ N .
It is neutral if λ(X) = λ(1 − X) for all X ∈ {0, 1}N . Note that if λ is neutral, then
Pi = 12 for all i ∈ N.

n
N
A random electorate λ is independent if for all X ∈ {0, 1} , λ (X) = (Pi )Xi (1 − Pi )1−Xi .
i=1
The Banzhaf random electorate is the unique neutral and independent random electorate
λ defined by λ (X) = 21n for all X ∈ {0, 1}N . In the social choice literature this model is
referred to as Impartial Culture (IC).
The Shapley-Shubik random electorate λ is defined by λ (X) = (n+1) 1
for all X ∈ {0, 1}N
(nk)
such that | {i ∈ N : Xi = 1} |= k. In the social choice literature, this model is referred to as
Impartial Anonymous Culture (IAC).5
A voting mechanism is a monotonic mapping C from {0, 1}N into {0, 1}.6
A mechanism C is anonymous if for all X ∈ {0, 1}N and all permutation σ over N ,
 
C(X) = C(X σ ) where X σ is the vector Xσ(1) , ..., Xσ(n) .
5
As shown by Straffin (1978), the Shapley-Shubik random electorate is equivalently defined as the mixture
according to the uniform density over [0, 1] of uniform and independent random electorates. This model has
been pioneered independently in voting theory by Chamberlain and Rothschild (1981), Fishburn and Gehrlein
(1976), Good and Mayer (1975) and Kuga and Nagatani (1974). His connection with the Shapley-Shubik
power index, i.e. when vote takes place between two alternatives, was discovered by Straffin (1978).
6
It is equivalently described by the inverse image C−1 (1) of 1 by C, i.e. the list W of coalitions such that
if {i ∈ N : Xi = 1} ∈ W , then C(X) = 1. The set W is refered to as the set of winning coalitions. Such a
mechanism is also called a simple game (Shapley (1962), Taylor and Zwicker (1999)).

3
 A mechanism C is neutral if for all X ∈ {0, 1}N , C(1 − X) = 1 − C(X). If λ is neutral
and C is neutral then λ (C) ≡ λ [X : C(X) = 1] = 12 .
Among the neutral voting mechanisms, we will consider Dic and Maj: Dici (X) = Xi
(individual i dictates his choice) and the ordinary majority mechanism Maj defined7 by:
  n
1 if 1≤i≤n Xi > 2 ,
Maj(X) =  n
0 if 1≤i≤n Xi < 2 .

Given a random electorate λ and a voting mechanism C, the influence (power) of indi-
vidual i ∈ N is defined as the probability that i is pivotal under mechanism C, that is:

Inf luence (i, λ, C) = λ[X : C(X−i ∪ {0}) = 0 and C(X−i ∪ {1}) = 1], (1)

and the expected utility of individual i, given our assumption that the utility can only take
two values: 1 for the best alternative and 0 for the worst, is the probability that the collective
decision coincides with individual i’s preferred alternative; it writes as:

Utility (i, λ, C) = λ[C(X) = Xi ]. (2)

On the aggregate side, the sensitivity8 (total influence) of the mechanism C is defined as:

T otal Inf luence (λ, C) = Inf luence (i, λ, C),
i∈N

and the expected aggregate utility writes as:


T otal U tility (λ, C) = Utility (i, λ, C).
i∈N

For any given random electorate λ, these two numerical evaluations define two orderings

Tλ I and
Tλ U over the set of mechanisms.
Note that since T otal U tility (λ, C) coincides with the average number of people who
agree with the collective decision, for all λ, the best mechanism according to
Tλ U is the
ordinary majority mechanism. This means that an utilitarian mechanism designer facing no
constraints in the choice of a mechanism will pick up the ordinary majority mechanism.
Obtained when λ is the Banzhaf electorate (the IC case), the celebrated Penrose formula
(1946) writes as9 :
1 1
U tility (i, IC, C) = + Inf luence (i, IC, C). (3)
2 2
7
When n is odd.
8
Since the word sensitivity can receive many different interpretations, we prefer to use the word total
influence which is used in the mathematical literature on Boolean functions.
9
A proof for this formula will be provided in section 3 when we establish a generalized Penrose formula.

4
 The Penrose formula implies that if λ is IC then
Tλ I =
Tλ U .
In the next section, we demonstrate a generalized version of the Penrose’s formula which
show that the affine relationship between influence and utility does not hold in general when
the electorate is not IC. A third term, that we call a correction term, appears in the
relationship.
Before turning to this general formula describing the link between influence and utility
when we allow for some correlations in the electorate, let us first note that there are neutral
and uniform random electorates λ such that the strict components of the orderings
Tλ I and

Tλ U disagree. As a very simple, motivating example, consider the case where n = 7 and λ
is the random electorate where the mass is distributed uniformly on all the profiles where 5
voters are on one side and 2 voters are on the other side.
Consider first the ordinary majority mechanism. The probability that an individual is
pivotal is equal to 0. Therefore, the total influence of this mechanism is equal to 0. On the
other hand, the social utility attached to each profile is equal to 5. There the expected total
utility is equal to 5.
Consider now the mechanism where 1 dictates his choice. The total influence of this
mechanism is equal to 1. The ex post social utility is either equal to 5 if 1 is on the majority
side or to 2 if 1 is on the minority side. Therefore, the expected social utility is equal to
5
7
× 5 + 27 × 2 = 29
7
. Without surprise given our observation that for all λ, the best mechanism
according to
Tλ U is the ordinary majority mechanism, we have 29 7
< 5.
With this random electorate λ, we have Dic1 ≻λ Maj and Maj ≻Tλ U Dic1 .
TI

Remark: We will focus in this note on the most popular two criteria used to evaluate
mechanisms: sensitivity and social utility/utilitarianism. Felsenthal and Machover (1999)
add to these two notions, the notion of majority deficit. It is defined as follows.10 Given a
profile of preferences X and a voting mechanism C, define the majority deficit of C at X as
the difference between the size of the majority group at X and the size of the winning group
if this difference is positive and 0 otherwise. This number measures the gap, if any, between
the majority and the minority. Given a random electorate λ , the majority deficit of C is the
expected value (over X) of the majority deficit of C at X.
Note that if λ is IC, then11 , the majority deficit of C is equal to:
T otal Inf luence (IC, Maj) − T otal Inf luence(IC, C)
,
2
and in that case, all three criteria agree on the ranking of mechanisms.
10
See e.g. page 60 in Felsenthal and Machover (1998).
11
See theorem 3.3.17 in Felsenthal and Machover (1998).

5
3 A Generalized Penrose Formula
In this section, we propose a generalization of the Penrose formula (3) to any random elec-
torate λ.
By definition (2), U tility (i, λ, C) = λ[C(X) = Xi ]. We use the fact that the two sets
{X : C(X−i ∪ {1 − Xi }) = 1 − Xi } and {X : C(X−i ∪ {1 − Xi }) = Xi } form a partition of
the state space, to decompose this expected utility as follows:

Utility (i, λ, C) = λ [X : C(X) = Xi and C(X−i ∪ {1 − Xi }) = 1 − Xi ]

+λ [X : C(X) = Xi and C(X−i ∪ {1 − Xi }) = Xi ] . (4)

Note that the first term on the right hand side of (4) is Inf luence (i, λ, C) (see definition
(1)).12 Besides, since the voting mechanism C is monotonic, if C(X−i ∪ {1 − Xi }) = Xi , then
C(X) = Xi , and therefore the second term on the right hand side of (4) can be simplified as:

λ [X : C(X) = Xi and C(X−i ∪ {1 − Xi }) = Xi ] = λ [X : C(X−i ∪ {1 − Xi }) = Xi ] .

Substituting in (4), one gets:

U tility (i, λ, C) = Influence (i, λ, C) + λ[X : C(X−i ∪ {1 − Xi }) = Xi ]. (5)

To get a formula closer to Penrose’s formula, note that:

Utility (i, λ, C) = 1 − λ[X : C(X) = 1 − Xi ]. (6)

Summing (5) and (6) and dividing by 2, one gets our generalized Penrose formula:
 
1 1 1 λ[X : C(X−i ∪ {1 − Xi }) = Xi ]
Utility (i, λ, C) = + Influence (i, λ, C) + .
2 2 2 −λ[X : C(X−i ∪ {Xi }) = 1 − Xi ]
(7)

Formula (7) is a generalization of Penrose’s formula (3) to any random electorate. Note
that the third term of the right hand side of (7), which can been seen as a correction term
12
Indeed,

λ [X : C(X) = Xi and C(X−i ∪ {1 − Xi }) = 1 − Xi ]

= λ [Xi = 0] . λ [X : C(X) = Xi and C(X−i ∪ {1 − Xi }) = 1 − Xi | Xi = 0]
+λ [Xi = 1] . λ [X : C(X) = Xi and C(X−i ∪ {1 − Xi }) = 1 − Xi | Xi = 1]
= (λ [Xi = 0] + +λ [Xi = 1]) . λ [X : C(X−i ∪ {0}) = 0 and C(X−i ∪ 1) = 1]
= λ [X : C(X−i ∪ {0}) = 0 and C(X−i ∪ 1) = 1] .

6
compared to the Penrose formula, receives a simple interpretation. It is (half) the difference
between the probability that the collective decision coincides with individual i’s wish even if
i were to vote against his interest, and the probability that the collective decision does not
coincide with individual i’s wish when i votes for his preferred alternative.

To check that we get Penrose’s formula when λ is the IC electorate, note that the
correction term on the right hand side of (7) can write as:
1
(λ[X : C(X−i ∪ {1 − Xi }) = Xi ] − λ[X : C(X−i ∪ {Xi }) = 1 − Xi ])
2
1
= λ[Xi = 1] ∗ (λ[X : C(X−i ∪ {0}) = 1 | Xi = 1] − λ[X : C(X−i ∪ {1}) = 0 | Xi = 1])
2
1
+ λ[Xi = 0] ∗ (λ[X : C(X−i ∪ {1}) = 0 | Xi = 0] − λ[X : C(X−i ∪ {0}) = 1 | Xi = 0]) .
2
When λ is independent, one can drop all the conditioning in the equality above, and the
correction term is:
1
Pi − ∗ (λ[X : C(X−i ∪ {0}) = 1] − λ[X : C(X−i ∪ {1}) = 0]) .
2

When λ is both independent and neutral (remember that neutrality implies that Pi = 1/2),
this term is equal to zero and formula (7) boils down to Penrose’s formula (3).

4 An application to the Shapley-Shubik IAC Random

Electorate
In this section, we apply our generalized formula to the popular Shapley-Shubik IAC random
electorate. In the first subsection we derive a general formula for the calculation of U tility
(i, IAC, C) for any voting mechanism C. In a second subsection, we use this formula to
compute this expected utility in the case of an anonymous mechanism. In a last subsection,
we compare the Banzhaf IC setting and the Shapley-Shubik IAC setting, highlighting the
differences between the two models.

4.1 A general formula for the IAC case

Consider our generalized Penrose formula (7). The term Inf luence (i, λ, C) is the probability
that i is pivot under mechanism C. When λ is the IAC electorate, the probability that out
of the n − 1 individuals in N \{i}, some k predetermined voters vote for A is the same
1 1
independently of the identity of these k voters, and this probability is n (n−1)
. Denoting by
k

7
γi (k, C) the number of coalitions in N\{i} with k votes for A for which i is pivotal, Inf luence
(i, IAC, C) writes as:
k=n−1
 γi (k, C)
Inf luence (i, IAC, C) =  . (8)
k=0
n n−1k

Consider now the correction term 12 λ[X : C(X−i ∪{1 − Xi }) = Xi ]− 12 λ[X : C(X) = 1−Xi ]
in (7), it is equal to:

1 λ[X : Xi = 0 and C(X−i ∪ {1}) = 0]+ 1 λ[X : Xi = 0 and C(X−i ∪ {0}) = 1]+
− ,
2 λ[X : Xi = 1 and C(X−i ∪ {0}) = 1 ] 2 λ[X : Xi = 1 and C(X−i ∪ {1}) = 0]

that is:
1 λ[X : C(X−i ∪ {0}) = 1 and Xi = 1]− 1 λ[X : C(X−i ∪ {1}) = 0 and Xi = 0]−
+ .
2 λ[X : C(X−i ∪ {0}) = 1 and Xi = 0] 2 λ[X : C(X−i ∪ {1}) = 0 and Xi = 1]

Denoting by αi (k, C) the number of coalitions in N\{i} with k votes for A, for which A
wins even if i votes B, and by βi (k, C) the number of coalitions in N\{i} with k votes for A,
for which B wins even if i votes A, and using the property of the IAC electorate, one gets
that the correction term is equal to:
k=n−1 k=n−1
1  1 1 1 1  1 1 1
αi (k, C)  n
 − n + βi (k, C) n −  n  .
2 k=0 n+1 k+1 k
2 k=0 n+1 k k+1

Since
1 1 1 1
 n
 − n = n−1 [2k − (n − 1)] ,
k+1 k
n k
when λ = IAC, one gets for the correction term:
1 1
λ[X :C(X−i ∪ {1 − Xi }) = Xi ] − λ[X : C(X) = 1 − Xi ]
2 2
k=n−1

1 αi (k, C) − βi (k, C) 2k − (n − 1)
=   × . (9)
2 k=0 n n−1
k
n+1

4.2 A focus on anonymous mechanisms in the IAC case

For the sake of illustration, let us focus on anonymous mechanisms13 . Consider an arbitrary
integer quota q ∈ [0, n + 1], such that option A is chosen if and only if at least q individuals
vote for A. The case q = 0 (resp. q = n + 1) corresponds to the mechanism which always
13
The formulas are much more cumbersome for non anonymous mechanisms.

8
selects option A (resp. option B). Let us first compute, for any k ∈ [0, n − 1], αi (k, q),
βi (k, q) and γi (k, q). Since the mechanisms are anonymous, we can drop the i subscript.
Given quota q:
 n−1
if k = q, ..., n − 1 and q ≤ n − 1,
α(k, q) = k
0 otherwise,
 n−1
if k = 0, 1, ..., q − 2 and q ≥ 2,
β(k, q) = k
0 otherwise,
 n−1
if k = q − 1 and 1 ≤ q ≤ n,
γ(k, q) = k (10)
0 otherwise.

We deduce from (8) that:

 1
for 1 ≤ q ≤ n,
Influence (i, IAC, q) = n
0 for q ∈ {0, n + 1}.

Remark: Note that this provides a simple proof for the well known claim14 that, for any
non constant anonymous mechanism C, T otal Inf luence(IAC, C) = 1. In the IAC case, any
difference in Total Utility between mechanisms therefore stems from the correction term.
Let us now compute the correction term (9).
Consider first the case 2 ≤ q ≤ n − 1. One gets by substituting the values for α(k, q) and
β(k, q):
1 1
λ[X :C(X−i ∪ {1 − Xi }) = Xi ] − λ[X : C(X) = 1 − Xi ]
2 2
k=n−1 k=q−2
1 1  
= [2k − (n − 1)] − [2k − (n − 1)] .
2 n (n + 1) k=q k=0

Simple computations yield:

k=n−1 k=q−2
 
[2k − (n − 1)] = q(n − q) and [2k − (n − 1)] = − (q − 1) (n + 1 − q) ,
k=q k=0

therefore
1 1 1 (n + 1 − q) q
λ[X : C(X−i ∪ {1 − Xi }) = Xi ] − λ[X : C(X) = 1 − Xi ] = − + .
2 2 2n n (n + 1)
(11)
14
Indeed as already pointed out, Influence (i, IAC, C) is the Shapley imputation of player i in the simple
game attached to C. If the game is symmetric, then all the players receive the same payoff in the Shapley
solution (as it is a symmetric solution) and since the total payoff is 1 (as the game is not constant), the claim
follows.

9
 Consider now the case q = 1 (the mechanism selects option A as soon as at least one
voter votes for A) or the symmetric case q = n (the mechanism selects option B as soon as
at least one voter votes for B).
 
When q = 1, α(0, 1) = 0 and α(k, 1) = n−1
k
for all k ∈ [1, n − 1], and β(k, 1) = 0 for all
k ∈ [0, n − 1]. Therefore, substituting in (9), one gets:
1 1
λ[X :C(X−i ∪ {1 − Xi }) = Xi ] − λ[X : C(X) = 1 − Xi ]
2 2
k=n−1
1  2k − (n − 1) (n − 1)
= = ,
2 n (n + 1) 2n (n + 1)
k=1

and therefore formula (11) is still valid. The same results can be shown to hold when q = n,
 
by noting that in that case, α(k, n) = 0 for all k ∈ [0, n − 1] and β(k, n) = n−1 k
for all
k ∈ [0, n − 2], β(n − 1, n) = 0.
Consider last the case of a constant mechanism (q = 0 or q = n + 1). It is straightforward
to check directly that λ[X : C(X−i ∪ {1 − Xi }) = Xi ] − λ[X : C(X) = 1 − Xi ] = 0. Indeed,
when for example q = 0, λ[X : C(X−i ∪ {1 − Xi }) = Xi ] = λ[X : Xi = 1] = 12 and
λ[X : C(X) = 1 − Xi ] = λ[X : Xi = 0] = 12 .
To summarize, we therefore get for the correction term when λ = IAC:

1 λ[X : C(X−i ∪ {1 − Xi }) = Xi ] 1
− 2n + (n+1−q)q
n(n+1)
for 1 ≤ q ≤ n,
=
2 −λ[X : C(X) = 1 − Xi ] 0 for q ∈ {0, n + 1}.

Combining the pivot term and the correction term, one gets that for all q, 0 ≤ q ≤ n + 1:
1 (n + 1 − q) q
U tility (i, IAC, q) = + ,
2 n(n + 1)
where which can be decomposed for 1 ≤ q ≤ n as
 
1 1 1 1 1 2 (n + 1 − q) q
U tility (i, IAC, q) = + + − + ,
2 2 n 2 n n n+1
where the second term in the sum on the right-hand side is the influence term and the third
term is the correction term; when q ∈ {0, n + 1}, both the influence and the correction term
are zero. This shows that all (non-constant) anonymous mechanisms yield the same power
( n1 ), but that yield different expected utilities.
Denoting q as ⌈θn⌉ where θ ∈ ]0, 1[ and ⌈x⌉ is the smallest integer strictly greater than
x, the correction term is approximately, for a large electorate:
 
1 1 (n + 1 − θn) θn 1 1 θ 1 1
− + ≃− + θ (1 − θ) + = θ(1 − θ) + θ2 − .
2n n (n + 1) 2 n n 2 n

10
 1
Note that for a large electorate, the correction term is much larger than the pivot term ( 2n ),
since the order of magnitude is θ(1 − θ). For a large electorate:

1 θ2
U tility (i, IAC, q) ≃ + θ(1 − θ) + where q = ⌈θn⌉. (12)
2 n
While simple, let us highlight the fact that:

θ2
U tility (i, IAC, q) − Utility (i, IAC, Pure Randomization) ≃ θ(1 − θ) + where q = ⌈θn⌉,
n
i.e., that the per capita gain of optimization compared to randomization is 0(1).

4.3 Comparison with the IC setting

By definition, under IC the correction term is absent. Using the same notations as in the
above subsections, and using the fact that when λ is the IC electorate, the probability that
out of the n − 1 individuals in N\{i}, some k predetermined voters vote for A is the same
1
independently of the identity of these k voters, and this probability is 2n−1 , we obtain that
for any voting mechanism C:
k=n−1
 γi (k, C)
Inf luence (i, IC, C) = .
k=0
2n−1

Not much can be said on the probability of being pivotal at this level of generality besides
the following inequality known as the edge-isoperimetric inequality15 :

1
T otal Inf luence (IC, C) ≥ 2λ (C) Log2 ,
λ (C)

where we have defined λ (C) as λ (C) = λ [X : C(X) = 1] (see section 2). If the mechanism is
neutral, then λ (C) = 12 and the edge-isoperimetric inequality implies that T otal Inf luence
(IC, C) ≥ 1. Equality holds when C = Dic.
We are not aware of any characterization of the influence ranking
Tλ I when λ = IC. Let
us focus on the subclass of anonymous mechanisms, and still denote by q the quota above
which option A is elected. Given the values for γi (k, q) (see (10)), we obtain that:

1 n−1
Inf luence (i, IC, q) = for q ∈
/ {0, n + 1}.
2n−1 q−1

If we assume that n and q are large, then from Stirling’s formula, we derive:
15
See Kalai and Srafa (2006).

11
 
1 1 n−1 (n − 1)n
Inf luence (i, IC, q) ≃ n−1 √ .
2 2π (q − 1)(n − q) (q − 1)q−1 (n − q)n−q
Denoting q as ⌈θn⌉ where θ ∈ ]0, 1[ and ⌈x⌉ is the smallest integer strictly greater than x,
the above formula writes as:

1 1 1
Influence (i, IC, q) ≃ √ .
2πn θ(1 − θ) (2θ (1 − θ)1−θ )n
θ

 
When θ = 12 , (2θθ (1−θ)
1
1−θ )n behaves as e
−γn
, where γ = ln 2θ θ (1 − θ)1−θ is a positive constant
(since the function θ ln θ + (1 − θ) ln(1 − θ) has a maximal value of − ln 2 (on the interval
[0, 1] ) at θ = 12 ). In contrast, when θ = 12 , we obtain:
 
n 2
Influence i, IC, ≃ .
2 πn
From Penrose’s formula, we deduce:
 
 1+ 1
if q = ⌈ n2 ⌉,
2 2πn
U tility (i, IC, q) ≃  (13)
 1 + √1 1
e−γn if q = ⌈θn⌉ with θ = 12 .
2 2πn θ(1−θ)

It is important to note from (13) that if λ is IC, then the social improvement over pure
randomization of any non constant anonymous mechanism q, T otal U tility (IC, q) − T otal
√
U tility (IC,Pure Randomization), is 0( n).
More generally, take any voting mechanism C. Using the fact that for all λ, the best
mechanism according to
Tλ U is the ordinary majority mechanism Maj, one gets that

T otal U tility (IC, C) − T otal U tility (IC, Pure Randomization)

≤ T otal U tility (IC, Maj) − T otal U tility (IC, Pure Randomization)
√
and for any mechanisms C, the social improvement over pure randomization is 0( n): the
per capita gain of optimization compared to randomization is of second order with respect
to the size of the population.
We have seen in the previous subsection that the picture looks very different when the
random electorate is not IC: in particular, if λ is IAC, then there are mechanisms C such
that the per capita improvement over pure randomization is 0(1).
In an environment with correlation, the mechanism design exercise is useful as it has a
first order effect when compared to pure randomization. In such case, any departure from
the optimal mechanism has an impact. In contrast, in the IC case, it has only second order
effects.

12
5 Concluding Remarks
This note has offered some insights on the differences between two popular social objective
functions in the area of voting mechanism design: T otal Inf luence and T otal U tility. In
the case of the IC model, the Penrose formula shows that the two social objective functions
coincide. While conclusions obtained with the IC model are insightful, this note show that
we should be careful about evaluating the qualities of alternative social mechanisms and
deriving the optimal one on the sole basis of the IC model. Indeed, with the Penrose formula,
we have seen that in the IC model, the exercise amounts to comparing second order effects
among themselves and that the gain with respect to randomness is questionable. Second, and
more importantly, we have shown that as soon as we introduce correlations across individual
preferences, the correction term dominates the pivot term in the evaluation of social utility.
Of course many important questions remains to be solved.
First, can we derive a full characterization of the two social objectives attached to the
orderings
Tλ I and
Tλ U for a large class of random electorates λ? Is it easy to recognize for
any pair of mechanisms and any of these two orderings which mechanism in the pair is the
best one?
Second, what happens if we depart from the assumption that the individual utilities can
only take two values, assumed to be the same across all individuals. In real world appli-
cations where, often, the players are not individuals but population of individuals (regions,
countries,...) this assumption is not relevant and should be replaced by other assumptions.
This is what is done in Barbera and Jackson (2004) and Beisbart, Bovens and Hartman
(2005). In such a context, it is no longer true that the best mechanism according to
Tλ U is
the ordinary majority mechanism. The characterization of the top element of
Tλ U for a large
class of independent random electorates is solved in Barbera and Jackson, and Beisbart et
al..
Third, optimization over C may be subject to constraints. The designer task is then to
find the best mechanism (according to any of the two objectives studied above), under some
set of constraints. One classical set of constraints consists in imposing to the mechanism to
be indirect or two-tier on the basis of a given partition of the individuals into areas (countries,
regions, districts,..): this means that the mechanism defines first for each of this group a
representative or temporary winner (A or B) for the group based upon the preferences in
the group and then a second mechanism to map the profile of temporary group winners into
the ultimate choice. In this literature, the first mechanism is often postulated (for instance
it is assumed to be the ordinary majority mechanism) and the designer only optimizes with

13
respect to the second mechanism. The contributions of Barbera and Jackson (2004) and
Beisbart, Bovens and Hartman (2005) can also be interpreted in such way: derivation of the
optimal second best mechanism. Given the second best nature of the problem, the ordinary
majority mechanism (direct democracy) is not a feasible choice.16

6 References
Banzhaf, J.F. (1965) ”Weighted Voting does not Work : A Mathematical Analysis”, Rutgers
Law Review, 19, 317-343.
Barbera, S. and M.O. Jackson (2006) ”On the Weights of Nations : Assigning Voting
Weights in a Heterogeneous Union”, Journal of Political Economy, 114, 317-339.
Beisbart, C. and L. Bovens (2013) ”Minimizing the threat of a positive majority deficit
in two-tier voting systems with equipopulous units”, Public Choice, 145, 75-94.
Beisbart, C., Bovens L, and S. Hartmann (2005) “A Utilitarian Assessment of Alternative
Decision Rules in the Council of Ministers”, European Union Politics, 6, 395—418.
Brams, S.J. and M. Lake (1978) ”Power and Satisfaction in a Representative Democracy”
in Game Theory and Political Science, Ordeshook (Ed), New York.
Chamberlain, G. and M. Rothschild (1981) “A Note on the Probability of Casting a
Decisive Vote”, Journal of Economic Theory, 25,152-162.
Curtis, R.B. (1972) ”Decision Rules and Collective Values in Constitutional Choice” in
Probability Models of Collective Decision Making, Niemi, R. and H. Weisberg (Eds), Merrill,
Columbus, Ohio.
Dubey, P. and L.S. Shapley (1979) ”Mathematical Properties of the Banzhaf Power In-
dex”, Mathematics of Operations Research, 4, 99-131.
Feix, M., Lepelley, D., Merlin, V. and J.L. Rouet (2004) ”The probability of conflicts in
a U.S. presidential type election,” Economic Theory, 23, 227-257.
Felsenthal, D.S. and M. Machover (1998) The Measurement of Voting Power, Edward
Edgar, Northampton.
Felsenthal, D. S. and Machover, M. (1999) Minimizing the mean majority deficit : the
second square-root rule. Mathematical Social Sciences, 37, 25-37.
16
This literature contains several results under IC. When the utilities takes two values, it has been
demonstrated by Felsentahl and Machover (1999) that the maximal
Tλ U mechanism is a weighted majority
mechanism where each group receives a weight proportional to the square root of its population. Then, if
groups are equipopulous, the optimal second tier mechanism is the ordinary majority mechanism. It has
been further demonstrated by Beisbart and Bovens (2013) that when the groups are equipopulous, T otal
U tility (λ, C) where C is the majority two-tier mechanism reaches it minimum when the partition is such
that the population in each group is about the square root of the total population.

14
 Fishburn, P.C. and Gehrlein, W.V. (1976) “Borda’s Rule, Positional Voting, and Con-
dorcet’s Simple Majority Principle”, Public Choice, 28, 79-88.
Good, I.J. and L.S. Mayer (1975) “Estimating the Efficacy of a Vote”, Behavioral Science,
20, 25 -33.
Kalai, G. and S. Safra (2006) ”Threshold Phenomena and Influence with Some Perspec-
tives from Mathematics, Computer Science, and Economics”, Computational Complexity and
Statistical Physics, Percus, A., Istrate G., and [Link] (Eds), Santa Fe Institute Studies on
the Sciences of Complexity, Oxford University Press.
Kuga, K. and H. Nagatani (1974) “Voter Antagonism and the Paradox of Voting”, Econo-
metrica, 42, 1045-1067.
Myerson, R.B.(1998) “Population Uncertainty and Poisson Games”, International Jour-
nal of Game Theory, 27, 375-392
Myerson, R.B.(2000) “Large Poisson Games”, Journal of Economic Theory, 94, 7-45.
Penrose, L.S. (1946) ”The Elementary Statistics of Majority Voting”, Journal of the
Royal Statistical Society, 109, 53-57.
Rae, D.W. (1969) ”Decision Rules and Individual Values in Constitutional Choice”,
American Political Science Review, 63, 40-56.
Shapley, L.S. (1962) ”Simple Games: An Outline of the Descriptive Theory”, Behavioral
Science, 7, 59-66.
Shapley, L.S. and M. Shubik (1954) ”A Method for Evaluating the Distribution of Power
in a Committee System”, American Political Science Review, 48, 787-792.
Straffin, P.D. (1978) ”Probability Models for Power Indices” in Game Theory and Political
Science, Ordeshook (Ed), New York.
Taylor, A.D. and W.S. Zwicker (1999) Simple Games, Princeton University Press, New
Jersey.
Weber, R.J (1995) “Approval Voting”, Journal of Economic Perspectives, 9, 39-49.
Wilson, M.C. and G. Pritchard (2007) “Probability Calculations under the IAC Hypoth-
esis”, Mathematical Social Sciences, 54, 244-256.

15