Available online at www.sciencedirect.

com

ScienceDirect
Journal of Economic Theory 160 (2015) 175–187
www.elsevier.com/locate/jet

Notes

Dynamically consistent voting rules

Madhav Chandrasekher
Department of Economics, Arizona State University, Main Campus, PO Box 879801, Tempe, AZ 85287-9801,
United States
Received 17 January 2014; final version received 18 August 2015; accepted 30 August 2015
Available online 5 September 2015

Abstract
This paper studies preference aggregation in a dynamic choice context. Voters face menus of options in
stages and the source of dynamics is that menus possibly get smaller across each stage. We call the family
of stage voting rules “dynamic voting rules” and provide an axiomatic characterization, on the domain of
single-peaked preferences, of dynamic voting rules that are strategy-proof and satisfy a second property,
inspired from choice theory, which we call dynamic consistency.
© 2015 Elsevier Inc. All rights reserved.

JEL classification: D03; D60; D70; D71; D81; D90

Keywords: Aggregation of preferences; Dynamic consistency; Strategy-proofness

1. Introduction

This paper conducts an axiomatic analysis of dynamic voting rules. To illustrate the sort of
choice problem we have in mind, imagine a firm has an open position and that it fills the position
by a vote of the partners. First, the firm solicits applications and forms a candidate pool, on
which it votes. When the firm makes an offer to its first choice, the candidate may not accept
the job, and candidates may exit the pool (e.g., because they have taken a position at another
firm). When a candidate exits the pool the firm faces a smaller pool after the first offer, and
subsequently smaller pools after subsequent offers. This is an example of a dynamic aggregate
choice problem. Decision nodes are indexed by the currently available pool of candidates and at

E-mail address: madhav.chandrasekher@asu.edu.

http://dx.doi.org/10.1016/j.jet.2015.08.012
0022-0531/© 2015 Elsevier Inc. All rights reserved.
176 M. Chandrasekher / Journal of Economic Theory 160 (2015) 175–187

each node committee votes are aggregated into an aggregate decision (e.g., making an offer to a
candidate). The rule that governs how committee votes are used to form an aggregate decision
(e.g., majority rule, super-majority rule, unanimity rule) is called the voting rule and, in theory,
each decision node could be governed by a different voting rule.
Two conditions will restrict the collection of voting rules across decision nodes. First, they
must induce truthful reporting of preferences at every decision node. This requirement is common
in the voting literature and is referred to as “strategy-proofness”. Strategy-proofness formalizes
the idea that voting rules should be non-manipulable. The second condition is that the voting
rules used to reach an aggregate decision at each node must be consistent with one another (e.g.,
the firm always makes the offer to the best qualified candidate in each round). We refer to this
second condition as “dynamic consistency”. The concept of dynamic consistency is borrowed
from decision theory. To explain the idea, consider a choice problem under uncertainty where
the uncertainty resolves in stages and, at each stage, a decision-maker’s (interim) preferences
adapt to the new information. Dynamic consistency disciplines the collection of interim prefer-
ences in the following manner (see, e.g., Siniscalchi, 2011): if one choice is interim preferred to
another no matter which way uncertainty resolves, then this preference is maintained on the node
immediately prior to when this uncertainty resolves. In our decision problem there is no uncer-
tainty. Nonetheless, we can extrapolate from this a general principle of dynamic consistency: if
there is no payoff relevant change in the choice environment between two successive decision
nodes, then whenever one choice is preferred to another at a successor node this preference is
maintained at its predecessor node.
We use this principle to define dynamic consistency in our context. The time-dependent ob-
jects in our setting are the choice pools (menus). Dynamic consistency will specify a condition
under which the aggregate choice from a larger menu, say B, is the same as the subsequent choice
from some menu A, where A is contained in B. Our condition is a formal generalization of Sen’s
criterion α. The application Sen had in mind when he formulated his axiom (see Sen, 1969) was
the problem of rationality of social choice functions (the problem being non-existence of transi-
tive social preferences, after Arrow). To describe Sen’s α, imagine option pools are considered in
sequence as choices are removed from an initial set. Sen’s criterion states that if the choice from
some initial set B is still present in a subsequent (and smaller) option set A, then the same choice
is made from A. We generalize this by requiring the aggregate choice to agree across two points
in time so long as the original choice is still available and each voter’s first-best alternative from
the larger option set is still present in the smaller option set. This is our definition of dynamic
consistency.
The normative justification for dynamic consistency is that it links sequential decision-making
to the idea of rational choice, i.e., choices are derived from preference maximization. Importantly,
in the single-agent setting dynamic consistency adds “some” rationality to the decision-making
process without forcing choices to come from preference maximization. In the choice under un-
certainty example, there is a different preference relation at each decision node and the supports
of these relations are distinct, so it wouldn’t make sense1 for there to be a single, over-arching
preference relation that is maximized at each node. In the sequential setting of Sen’s α, requiring
choice to be rational (preference maximization) makes sense but it also frustrates Sen’s goal of
studying problems in which the “context” (the option set itself) influences a decision-maker’s
preferences. When we shift to the aggregate, however, the question of rationality – equivalently,

1 Excepting uninteresting cases, e.g., all acts (consumption plans) are indifferent or uncertainty is trivial.
 M. Chandrasekher / Journal of Economic Theory 160 (2015) 175–187 177

whether there is a representative agent – is both well-defined and does not cut against the ques-
tion of dynamic consistency. We thus ask two questions here. First, assuming strategy-proofness,
what does the class of dynamically consistent voting rules look like? Second, what does the
sub-class of rational voting rules look like?
The main result presents a pair of equivalences. First, we consider dynamic voting rules that
are strategy-proof and dynamically consistent. We characterize the set of such rules by proving
a representation theorem. Next, we consider rules that are strategy-proof and rational – so that
they admit a representative agent – and find that the class of such rules coincides with those
that are strategy-proof and dynamically consistent. Generally, rules that admit a representative
agent are “more rational” than those that are merely dynamically consistent, so they are certainly
nested within the class of rules described in the representation theorem. The surprise is that
the converse also holds: starting with any strategy-proof and dynamically consistent rule, we
show by construction that this rule must admit a representative agent. Since dynamic consistency
is strictly more general than rationality, without any other assumptions there would be a gap
between these two classes of rules.2 When we impose strategy-proofness, this gap vanishes;
therefore, once we require a strategy-proof dynamic voting rule to exhibit a little rationality (i.e.,
dynamic consistency) we force it to be fully rational.

Related literature. The contribution of this paper is that it studies, from an axiomatic viewpoint,
the joint implications of dynamic consistency and strategy-proofness when the voting problem
is itself dynamic. Dynamic consistency has received attention in the single-agent setting, e.g.,
see Machina (1989) for a survey, but in the aggregate the issue has received comparatively less
attention from an axiomatic viewpoint. The analysis in Jackson and Yariv (forthcoming) is clos-
est in purpose to our paper. These authors analyze the problem of aggregating preferences over
(dated) consumption streams and prove that only dictatorial collective choice functions can be
both Pareto efficient and time-consistent. Collective dynamic consistency is studied via a dy-
namic games approach in Bernheim and Ray (1989) and Lagunoff (2009). Bernheim and Ray
(1989) model collective dynamic consistency via a refinement of the concept of renegotiation-
proofness of strategies. In Lagunoff (2009), the per-period voting rule is itself a strategic choice,
and dynamic consistency is defined in terms of the per-period optimality condition that deter-
mines the choice of rule. This is shown to be essentially equivalent to time-stability, i.e., rules
that are chosen in one period are also chosen in the next. Our work complements a rich literature
in static strategy-proof social choice. In a seminal piece, Moulin (1980) characterizes strategy-
proof voting rules on the domain of single-peaked preferences. The voting rules defined in this
paper are a dynamic version of the rules characterized by Moulin (1980). More precisely, we
use a reformulation of this rule found in Barberà (2001) (originally presented in Barberà et al.,
1997).

2. Axioms and model

Assume the issue space X = {x1 < x2 < · · · < xk } ⊆ R is finite and that there is a finite set, say
{1, 2, . . ., N}, of voters in the population. Each voter has a single-peaked, complete, transitive,

2 For a simple example, define the outcome to be the median of the set of voter peaks (i.e., not counting multiplicity)
when this set has odd cardinality and the k + 1’st peak when it has even cardinality 2k. This is dynamically consistent,
but not rational.
178 M. Chandrasekher / Journal of Economic Theory 160 (2015) 175–187

and strict preference relation, i , on X.3 A generic profile of voter preferences is denoted as ,
and the set of profiles of single-peaked preferences is denoted as R. For any subset A ⊆ X, let
the restricted profile of preferences be denoted as A , and the space of such profiles as RA .4 The
primitive in this paper is a system of social choice functions (SCF’s), {A }, indexed by menus,
where A : RA → A. There is implicitly a finite time-horizon, T , and the function A describes
the voting rule in place if and when the time t option set is the menu A. Since the time-dependent
objects are the menus, there is also an implicit transition process that determines what the next
period’s menu will be as a function of the current menu and the current aggregate choice. The
only requirement on this process is that the time t + 1 menu be a subset of the time t menu, but
it is otherwise not mentioned in our axioms; hence, we suppress it from our description of the
primitive. As is standard, we use   to denote reported preferences.

Definition 1 (Strategy-proofness). A system {A } is strategy-proof if for all  and for each A
i , −i ), ∀ 
we have A (i , −i ) i A ( i .

In other words, each member of the system {A } is a strategy-proof function. This will
(typically) be a necessary condition for (subgame perfect) implementability.5 It is, however,
not necessarily sufficient, since equilibrium notions used in the implementation literature (e.g.,
subgame perfection) imply period-by-period strategy-proofness, but need not be implied by it.
Henceforth, we omit notation to distinguish reported preferences from actual preferences. We
call a family of profiles constant if, when A ⊆ B, we have (B )|A = A .

Definition 2 (Rationality). A system of SCF’s {A } is rational if for any constant family {A }
there is a single order ∗ such that A (A ) = arg maxx∈A (∗ )|A .

Rationality requires that aggregate choices, {A }, be derived from maximization of a single
order,  ⊆ X × X. In other words, there is a representative agent. We now relax this criterion by
requiring the system {A } to be consistent in the sense of Sen’s α. Since this is general enough
to allow choices at distinct decision nodes to come from distinct preferences, this will be our
operative definition of dynamic consistency. For a profile of preferences B over candidates in
pool B, let {a1 (B ), . . ., aN (B )} denote the profile of peaks (equiv. bliss/ideal points) of each
individual preference i comprising the profile B .

Definition 3 (Dynamic consistency). Fix A ⊆ B. A system of SCF’s {A } is dynamically con-

sistent if whenever {a1 (B ), . . ., aN (B )} ⊆ A and B (B ) ∈ A, then B (B ) = A (A ).

While Sen’s α is similar to the definition, it is weaker than Sen’s α since it only imposes choice
consistency on nested pairs (A, B) for which all ideal points from B are present in A. We close

3 Recall that an order  is single-peaked and strict on X if there is some x ∗ = x ∈ X such that x ≺ x ≺ x =
i j 1 i 2 i j
x ∗ , x ∗ i xj +1 i xj +2 i · · · i xk . Placing elements of X left to right along the number line (w.r.t. the natural
∗
ordering on R) the preference i has a unique local and global maximum at x , so that any utility representing the
preference is hump-shaped on X with a peak at x ∗ .
4 Observe that if  is single-peaked and strict, then its restriction to A is also single-peaked and strict.
X
5 Since the horizon of the dynamic voting problem is finite, whatever the menu of options is in this last period (it need
not necessarily be singleton), subgame perfection would require that the SCF induces truthful reporting of preferences in
this period.
 M. Chandrasekher / Journal of Economic Theory 160 (2015) 175–187 179

with two standard restrictions: anonymity and unanimity. Both of these are static restrictions,
hence we suppress reference to the menu.

Definition 4 (Anonymity). For any domain of preference profiles, R, let  denote a generic
permutation of the identities of the N voters and let  := ((1) , (2) , . . ., (N ) ). Say that
 satisfies anonymity if () = ( ), ∀.

Definition 5 (Unanimity). Let  be a profile such that 1 = · · · = N , i.e., all voters have
identical preferences over outcomes. Then, () = ai ().

Anonymity is the requirement that the aggregate choice only depends on the set of voter
preferences and not on the identity of the voter who has a particular preference within this set.
Unanimity requires that the aggregate choice pick out the common, most-preferred outcome
when everyone has the same preference. Now turn to the voting model. The definition below uses
an object that we dub the “q-min”. The q-min is a special case of a quantile. Recall that if we have
a probability distribution P on a subset X ⊆ R, then for any q ∈ [0, 1] the q-tile w.r.t. P (when
defined) is the element xq ∈ X such that P (x ≤ xq ) = q. The q-min denotes the special case of
this where we take a uniform distribution over the profile of voter peaks {a1(), . . ., aN ()}.6

Definition 6. An SCF  : R → X is a monotone threshold rule if for each x ∈ X there is a

qx ∈ {1/N, 2/N, . . ., 1} such that
() = min max{qx - min{a1 (), . . ., aN ()}, x}
x∈X

where qx1 ≥ qx2 ≥ · · · ≥ qxk ≡ 1/N .7

The thresholds defining the rule admit a nice interpretation. Let x,b denote a two-peaked
profile where everyone’s peak is either x or b, and where b ( = x) is the maximum (w.r.t. ≤)
on X. Whenever  admits a monotone threshold representation, these thresholds are recovered
via:
N qx = min |{i : ai (x,b ) = x}|.
x,b :(x,b )=x

Thus, ignoring the pre-multiplication by N , the array of thresholds {qx } represent the (resp.) car-
dinalities of the minimal decisive coalitions for each issue x ∈ X. Barberà et al. (1997) interpret
“support for x” at a profile  to be {i : ai () ≤ x}, i.e., all voters whose bliss point is at or to the
left of x. We visualize the outcome of the monotone threshold rule with the following schematic
(see Fig. 1).
For this discussion, assume all voter peaks are distinct, and ask the question: given the profile
of preferences , is there enough support for any given x? The function g(x) denotes the “pivotal

6 Concretely, for any set {a , . . ., a } of N distinct elements the q-min{a , . . ., a } picks out the element in the set
1 N 1 N
{a1 , . . ., aN } at the N q’th N -tile, e.g., when all the peaks are distinct and q = 2/N , the q-min picks out the second lowest
voter peak, when q = 3/N the q-min picks out the third lowest, and so on. When some voters have the same peak we can
still formally define the q-min the same way – it just means that, for example, the 2/N -min might equal the 3/N -min.
How we label voters with the same peak does not affect the q-min value.
7 The value of the threshold q
xk in a monotone threshold representation is not typically pinned down, beyond the
requirement that qxk ≤ qxi . We put qxk ≡ 1/N since we consider a dynamic version of this rule and need to make a
consistent selection for qxk across choice problems.
180 M. Chandrasekher / Journal of Economic Theory 160 (2015) 175–187

Fig. 1. The monotone threshold rule.

voter” line and picks out the ideal point of the pivotal (absent ties) voter on issue x, for the given
profile . In terms of the picture, when the g(x) line lies above f (x) there is not enough support
for x and the pivotal voter’s preference wouldn’t matter because the support for x is well below
the threshold qx . However, the pivotal voter’s preference matters at the intersection point. For
this x, the support for x (at ) exactly equals the minimal support required to pass x. When x
is to the right of the fixed point, the qx -min is smaller than x. Hence, support for x exceeds the
minimal amount required to pass x. At the fixed point, say x ∗ , we have N qx ∗ = |{i : ai () ≤
x ∗ }|. Since the qx -min is less than x ∗ , strategy-proofness implies the aggregate choice cannot
be x; else, the N qx ∗ voters with peaks at or less than x ∗ could report a peak of x ∗ and pull the
aggregate choice closer to their (resp.) bliss points.8 Precisely which SCF’s admit a monotone
threshold representation?

Proposition 1. (See Moulin, 1980; Barberà et al., 1997.) An SCF  is anonymous, unanimous,
and strategy-proof if and only if it is a monotone threshold rule.

Moulin (1980) was the first to prove this result. In fact, the proposition is a corollary of the
generalized median rule representation in Moulin (1980), () = minS∈2N \∅ maxi∈S [ai (), aS ]
(where the power set of voters is denoted 2N and aS ∈ R). Barberà et al. (1997) study a conceptu-
ally distinct voting problem and along the way introduce a reformulation of Moulin’s representa-
tion in the language of left(right) decisive coalitions. Monotone threshold rules are precisely the
anonymous “left coalition systems” of Barberà et al. (1997). We use these as the building block
for the following class of dynamic voting rules. First, let’s introduce some notation:

• Let Aad denote a generic menu of the form A ∪ {x}, x ∈ X\A.9

• Let I(x) := {y ∈ X : qy = qx } denote the q-indifference class of x.
• Let xA+ := min{y ∈ A : y > x} (we suppress the subscript when A = X).

8 Two additional comments. First, this discussion presumes that there is necessarily a point of intersection between the
“lines” f (x) = x and g(x) = qx -min at the profile . Since X is discrete, there are profiles where these lines don’t have
an intersection. For such cases, we need to give a separate argument (see Appendix A), which we omit for this heuristic
discussion. Second, the definition of “support” for x privileges support for x coming from voters with ideal points to the
left of x over those with an ideal point to its right. Barberà et al. (1997) show that the SCF admits an equivalent “dual”
representation in terms of what they call “right” decisive coalitions. Here, support is measured by the set of voters with
ideal point at or to the right of x.
9 The “ad” is for adjacent.
 M. Chandrasekher / Journal of Economic Theory 160 (2015) 175–187 181

Definition 7. We say that a system {A } is a monotone threshold family if, for each set A, A is
a monotone threshold rule with thresholds {qxA }x∈A satisfying:

1. (MON – Monotonicity) qxA ≤ qxB , ∀x ∈ A, A ⊆ B with equality if (i) xA+ = xB+ or (ii) I(x) ∩
{y ∈ A : y > x} = ∅.
ad
2. (BD – Bounded Decreases) qxA ∈ [qx (A), qx ], where qx (A) := maxAad qxA+ .10

The intuition for both (MON) and (BD) relies on interpreting the integers N qxA as cardinalities
of the minimal decisive coalition. (MON) is based on the idea that as fewer options are available
there is less room for compromise. Ceteris paribus, this should increase the “power” of any
given decisive coalition, which suggests that the size of the minimal decisive coalition should get
smaller. When there is an increase in coalition power, (BD) places a bound on how much it can
ad
increase, i.e., qxA ≥ maxAad qxA+ . To explain this bound consider the size of the minimal decisive
coalition on the issue (say z) nearest and to the right of x that was present in A ∪ {z}, but has now
been removed. A consequence of dynamic consistency is that this must be a lower bound on the
size of the minimal decisive coalition for x (when the menu of options is A). No hypotheses are
imposed on paths leading to the current round menu A. Hence, any menu of the form A ∪ {x} is
an admissible predecessor to a current menu – which accounts for the “max” defining the lower
bound.

An example. Consider X = {x < y < w < z} and N = 3 (4 outcomes, 3 voters). For brevity,
we write q in place of N q. The monotone threshold family is defined by the sizes of minimal
decisive coalitions for each (non-singleton) choice problem:

1. X : (qx = 3, qy = 3, qw = 1, qz = 1).
2. X\x : (qy = 3, qw = 1, qz = 1), X\y : (qx = 3, qw = 1, qz = 1), X\w : (qx = 3, qy =
2, qz = 1), X\z : (qx = 3, qy = 3, qw = 1).
3. X\{x, y} : (qw = 1, qz = 1), X\{x, w} : (qy = 1, qz = 1), X\{x, z} : (qy = 3, qw = 1).
4. X\{y, z} : (qx = 3, qw = 1), X\{y, w} : (qx = 2, qz = 1), X\{z, w} : (qx = 3, qy = 1).

Note that there are dynamics in the coalition sizes, so that the stage voting rule is not the same
across rounds, but (MON) and (BD) together control how coalition power grows. For example,
when the minimal decisive coalition for x is smaller when the option set is X\{y, w} (compared
to when the option set is X), its size must be 2 and not 1, on account of (MON) and (BD).

3. Representation theorems

Theorem 1. An SCF system {A } satisfies anonymity, unanimity, strategy-proofness, and dy-
namic consistency if and only if it is a monotone threshold family.

We give a sketch of how (MON) and (BD) imply dynamic consistency, focusing on the special
case in which the fixed point schematic (Fig. 1) for the monotone threshold applies. The general

10 In the definition of q (A), rather than q Aad we should write q Aad . This notation is a little cumbersome, so we omit
x x+ x +ad
A
the subscript on x + with the understanding that we are applying the “plus” operation on the menu Aad . Also note that
when x = max{z ∈ A} we have qx (A) = 1/N , so that our convention on qx (i.e., qx ≡ 1/N ) is consistent with the bound.
182 M. Chandrasekher / Journal of Economic Theory 160 (2015) 175–187

proof (along with the sufficiency argument) is left to Appendix A. Fix A ⊆ B, where B = A ∪ {z}
and assume the choice in B is some y and that all voter peaks from B are present in A. By
monotonicity of thresholds, the only candidate for a switch would be some x < y. Decisive
coalitions for x gain power going from B to A, but (BD) controls the gain in power. If z (the
removed option) is immediately to the right of x, then the minimal decisive coalition for x (on
A) cannot be smaller than the minimal decisive coalition for z (on B). But since there was not
enough support for z (since y = B (B ) is a fixed point) when the option set was B, there cannot
be enough support for x when the option set is A. If we remove multiple elements to reach A
from B, we use (BD) to construct a chain, A0 = A, A1 := A ∪ {x 1 }, A2 = A1 ∪ {x 2 }, . . ., Ak =
Ak−1 ∪ {x k } := B, and apply this argument along each step of the chain.

Theorem 2. An SCF system {A } satisfies anonymity, unanimity, strategy-proofness, and ra-
tionality if and only if it is a monotone threshold family. Moreover, the representative agent’s
preferences are always single-peaked.

Proof of Theorem 2. Rationality clearly implies dynamic consistency; hence, it remains to con-
struct a representative agent starting with a monotone threshold family. Note that, if the family
were rational, the only candidate for a rationalizing preference is the binary relation on doubleton
choice problems.

Step 1: Constructing the (putative) representative agent’s preference R.

Fix a constant family {A }. Define R via xRy ⇔ {x,y} ({x,y} ) = x. Clearly, R is a com-
plete binary relation. We claim that R is not only transitive (and hence an order), but also
single-peaked.

Lemma 1. R is transitive.

The proof just involves a check of several cases, hence is left to Appendix A.

Lemma 2. R is single-peaked.

Proof of Lemma 2. Consider two putative local maxima, say x ∗ < x ∗∗ , and put x + (resp.
x − ) nearest to x ∗ from the right (resp. left).11 The hypothesis that x ∗ is a local maximum
implies x ∗ Rx − , x ∗ Rx + . Ditto for x ∗∗ , where we take points to the immediate left and right
of x ∗∗ . Note that x + = x ∗∗ . Since x ∗ Rx + the argument for Lemma 1, case (i) (see Ap-
pendix A) shows that x ∗ Rx ∗∗ as well. Hence, the (global) maximum of the order R must
be the minimum (w.r.t ≤) of the local maxima. Let x ∗ denote this value. We now verify that
the order R is single-peaked with peak at x ∗ , so there is in fact only one local maximum.
By choice of x ∗ , the order is strictly decreasing to the left of x ∗ (when x ∗ is minimal w.r.t
≤ in X this statement is vacuously true, as there is nothing to the left of it). Now check that
it is strictly decreasing to the right of x ∗ as well. Take any (y, z) with x ∗ < y < z. Consider
{x ∗ ,y}
A := {x ∗ , y, z}. Since qxA∗ = qx ∗ we have qxA∗ -min(a1 (A ), . . ., aN (A )) = x ∗ . Hence, by
{y,z}
monotonicity of thresholds, qyA -min(a1 (A ), . . ., aN (A )) = x ∗ . Since qyA = qy , we obtain:
{y,z}
qy -min(a1 ({y,z} ), . . ., aN ({y,z} )) = y, so that yRz. 2

11 We have introduced x − := max{y ∈ X : y < x}.

 M. Chandrasekher / Journal of Economic Theory 160 (2015) 175–187 183

Step 2: Showing that the aggregate choice function {A } maximizes the single-peaked order R.
Let xA denote the aggregate choice and denote the R-argmax via xRA . If (towards contra-
+ +
diction) xA < xRA , find x A
≤ xRA (in A) and note that x A
RxA . For brevity of notation,
+ +
put x := xA , x := xA . We claim qx -min(a1 (A ), . . ., aN (A )) ≤ x. Let z be an argument
A

on which A (A ) is attained, i.e., max{qzA -min(a1 (A ), . . ., aN (A )), z} = x = A (A ). If
z = x, the claim is obvious. If z < x, by monotonicity of thresholds we have qzA ≥ qxA . Since
qzA -min(a1 (A ), . . ., aN (A )) = x, we obtain qxA -min(a1 (A ), . . ., aN (A )) ≤ x. Now note that
{x,x + } {x,x + }
qx = qxA , which then implies that qx -min(a1 ({x,x + } ), . . ., aN ({x,x + } )) = x as well –
contradicting the hypothesis that x + Rx. Similarly, we obtain a contradiction if xA > xRA .
It follows that xA = xRA . 2

4. Conclusion

This paper studies the question of consistent preference aggregation in the context of dynamic
choice. Voters are sequentially presented with menus and menus are (possibly) shrinking over the
sequence. For each menu, there is a possibly distinct voting rule, and we characterize families of
voting rules (calling these “dynamic voting rules”) that satisfy strategy-proofness and a second
property dubbed “dynamic consistency”. The latter is based on the idea that a forward look-
ing decision-maker should make the same choice today that he/she would anticipate making in
the future were there no payoff-relevant change to his/her choice context. A class of rules that
satisfies both strategy-proofness and dynamic consistency is the set of rational rules – that is,
aggregation rules that admit a representative agent. We show that strategy-proof rules that sat-
isfy rationality are equivalent to those satisfying the prima facie more general notion of dynamic
consistency.

Acknowledgments

I thank the editor and anonymous referees for their feedback.

Appendix A

Proof of Lemma 1. This amounts to verifying six cases. In all cases, the premise is xRy & yRz,
and we want to show this implies that xRz. For brevity of notation, we move interchangeably
between q and N q – it will be clear from context when we mean the integer N q vs. the N -tile q.
Case (i): x < y < z.
We will show: If x < y < z and xRy, then xRz, ∀z > y. This stronger claim is also used
{x,y}
in Lemma 2. Consider the menu A := {x, y, z}. We must have qxA = qx by equality condi-
tion (i) defining (MON), denoted with the common value, q. It follows that q-min(a1 (A ), . . .,
{x,z}
aN (A )) = x. Since qx ≤ qxA , it follows that {x,z} ({x,z} ) = x, so that xRz.
Case (ii): x < z < y.
We claim this case is incompatible with the premise. Again, take A := {x, y, z}, and note that
{x,y} {z,y} {x,y} {z,y}
qx ≥ qzA (by (BD)) and qz = qzA (by equality (i) in (MON)). Hence, qx ≥ qz . Since
yRz, this implies that yRx as well, a contradiction.
Case (iii): z < x < y.
184 M. Chandrasekher / Journal of Economic Theory 160 (2015) 175–187

{z,x} {z,y}
Put A := {x, y, z}. Note that qzA = qz and qzA ≥ qz . The premise yRz implies
{z,y} {z,x}
qz -min(a1 ({z,y} ), . . ., aN ({z,y} )) = y. We claim this implies qz -min(a1 ({z,x} ), . . .,
aN ({z,x} )) = x. Note that any voter i with ai (A ) peaked at x has a restriction to {z, y} ei-
ther peaked at z or y. The fact that yRz implies that the sum total of those voters with A
peaked at z, i.e., |{i : ai (A ) = z}|, plus those with A peaked at x whose restriction to {z, y} is
{z,y}
peaked at z, i.e., |{i : ai (A ) = x, ai ((A )|{z,y} ) = z}|, is less than qz :
{z,y}
|{i : ai (A ) = z}| + |{i : ai (A ) = x, ai ((A )|{z,y} ) = z}| ≤ qz .
{z,y} {z,x} {z,x}
Since qzA ≥ qz and qzA = qz , we obtain qz -min(a1 ({z,x} ), . . ., aN ({z,x} )) = x. This
implies that xRz.
Case (iv): z < y < x.
{y,x} {z,y}
Putting A = {x, y, z}, the equality condition in (MON) implies qyA = qy , qzA = qz .
{z,x} {y,x} {z,x}
Note that qz ≥ qyA = qy (the inequality by (BD)). Hence, qz -min(a1 ({z,x} ), . . .,
aN ({z,x} )) = x, implying that xRz.
Case (v): y < x < z.
{y,x} {y,z} {x,z} {x,z}
Mirror arguments imply qyA = qy and qy ≥ qxA ≥ qx . This implies that qx -
min(a1 ({x,z} ), . . ., aN ({x,z} )) = x. Hence, xRz.
Case (vi): y < z < x.
{y,z}
This case is also incompatible with the premise. Note that it implies qyA = qy , which in turn
{y,x}
implies that qyA -min(a1 (A ), . . ., aN (A )) = y. Since qyA ≥ qy , this implies {y,x} ({y,x} )
= y, a contradiction. 2

Proof of Theorem 1. Necessity. The only non-trivial check is necessity of dynamic consistency
(DC). Put xA− := max{y ∈ A : y < x}.
Case 1: qxB -min(a1 (B ), . . ., aN (B )) = B (B ) = x.
This is the case where the social outcome is the “fixed point” depicted in Fig. 1. For any
xj > B (B ), we have max{qxAj -min(a1 (B ), . . ., aN (B )), xj } > B (B ). Moreover, for
xj = B (B ) we have, since qxAj ≤ qxBj , max{qxAj -min(a1 (A ), . . ., aN (A )), xj } = xj . It re-
mains to check that for xi < xj (= B (B )), we have
qxAi -min(a1 (A ), . . ., aN (A )) ≥ xj .
Find the closest xi to the left of xj for which xi ∈ A. Passing to the monotone threshold repre-
sentation on A, it suffices to verify qxAi -min(a1 (A ), . . ., aN (A )) ≥ xj . To this end, we claim
that qxAi ≥ qxBj .
If xi , xj are in the same q-indifference class, then qxAi = qxBi ≥ qxBj , where the equality follows
from equality condition (ii) defining (MON). If they are not in the same q-indifference class,
then let x 0 := xi , x 1 , . . ., x k be a path in B connecting xi to x k := xj− (from left to right). Put
A0 := A, A1 := A ∪ {x 1 }, A2 = A1 ∪ {x 2 }, . . ., Ak = Ak−1 ∪ {x k }. By property (BD) we have:
qxAi ≥ qxA11 ≥ qxA22 ≥ · · · ≥ qxAkk .

Since xj ∈ Ak , the equality condition (i) (defining (MON)) implies qxAkk = qxBk . Since x k = xj−
(in B), we have qxBk ≥ qxBj . It follows that qxAi ≥ qxBj . This proves A (A ) = B (B ) in this case.
 M. Chandrasekher / Journal of Economic Theory 160 (2015) 175–187 185

We now turn to the case where the value of B (B ) is not attained at the intersection of
the pivotal voter and default choice line precisely because, on account of the discreteness of the
choice space and the given profile of preferences, these two lines do not share a common point.
Case 2: qxBj -min(a1 (B ), . . ., aN (B )) = xj , for any point xj on which the value B is at-
tained, i.e., B (B ) = max{qxBj -min(a1 (B ), . . ., aN (B )), xj }.
Take any xj on which the maximum occurs. There are two possibilities: either (i) the max-
imum occurs on qxBj -min(a1 (B ), . . ., aN (B )), or (ii) the maximum occurs on xj . For some
intuition, we reproduce the schema from Fig. 1. The two cases correspond (resp.) to the minimal
policies on the thickened red locus of points: (For interpretation of the colors in this figure, the
reader is referred to the web version of this article.)

In case 2(i), the marked point would actually be lower (w.r.t. ≤) on the y-axis than the
marked point to the right, and in the other case, the marked point on the right is the minimal
of the two. Consider the former case, i.e., qxBj -min(a1 (B ), . . ., aN (B )) > xj . Since there is
at least one such xj , we will (wlog) take xj to denote the maximal xj on which the value
B (B ) is attained and for which qxBj -min(a1 (B ), . . ., aN (B )) > xj . As above, we first need
to check that for xi < xj we have qxAi ≥ qxBj . The argument for this does not use the premise
that qxBj -min(a1 (B ), . . ., aN (B )) > (<)xj and proceeds as in the preceding case (i.e., we
create a path from xi to xj− and use the threshold bounds (BD) to show qxAi ≥ qxBj ), so we
omit the repetition. Hence, we know that qxAi ≥ qxBj whenever xi < xj . Now take xj to be so
that B (B ) = max{qxBj -min(a1 (B ), . . ., aN (B )), xj } = qxj -min(a1 (B ), . . ., aN (B )) > xj
(and xj is the ≤-maximal such element). Consider xB+ nearest to xj from the right (in B) and
note that we must have
(∗) qxBj -min(a1 (B ), . . ., aN (B )) = xB+ .
Since qxBj -min(a1 (B ), . . ., aN (B )) = (B ) ∈ A by hypothesis, it follows that qxA+ -
B
min(a1 (A ), . . ., aN (A )) ≤ xB+ . If we have equality then A (A ) = B (B ). If the in-
equality is strict, then consider two possibilities (i) xj ∈ A or (ii) xj ∈
/ A. In the latter case,
since we assume B (B ) = xB+ ∈ A, A (A ) = max{qxA+ -min(a1 (A ), . . ., aN (A )), xB+ }, as
B
qxAi ≥ qxBj , ∀xi < xj . In the former case, we have (putting x := xj to omit a double subscript)
xB+ = xA+ . Hence, qxA = qxB , implying that A (A ) = B (B ) in this case.
Now case 2(ii), where xj > qxBj -min(a1 (B ), . . ., aN (B )) and xj = B (B ). It fol-
lows that for any x < xj we have qxB -min(a1 (B ), . . ., aN (B )) > x, and for any x > xj
186 M. Chandrasekher / Journal of Economic Theory 160 (2015) 175–187

we have x > qxB -min(a1 (B ), . . ., aN (B )). We check that for x < xj , qxA -min(a1 (A ), . . .,
aN (A )) ≥ xj . Let x − denote the element in B closest to xj from the left. Consider x < x − .
We claim that qxA ≥ qxB− . Arguing as above, we use the inequality bounds (BD) to find a path,
x = x 0 , x 1 , . . ., x k = x − , of elements (from left to right) in B leading from x to x − . Let A0 :=
A
A, A1 := A ∪ {x 1 }, A2 := A1 ∪ {x 2 }, . . ., Ak . By (BD), we have: qxA ≥ qxA11 ≥ qxA22 ≥ · · · ≥ qx −k .
Moreover, since xj = (x − )+ and xj ∈ Ak equality condition (i) implies qx −k = qxB− . Hence,
A

qxA ≥ qxB− . It follows that for any x < x − , x ∈ A we have

max{qxA -min(a1 (A ), . . ., aN (A )), x} ≥ qxA -min(a1 (A ), . . ., aN (A ))

≥ qxB− -min(a1 (A ), . . ., aN (A ))
> x−.
This implies that, if x − ∈
/ A, then the value of A (A ) is attained on xj – which, in turn, implies
A (A ) = B (B ) since qxAj ≤ qxBj , xj ∈ A, and xj > qxBj -min(a1 (B ), . . ., aN (B )). If x − ∈
A, then by construction, we have (setting x ≡ x − ) xA+ = xB+ so that qxA− = qxB− . Hence, A (A )
= B (B ) in this case as well. This concludes the proof of necessity of DC.
Sufficiency: Apply Proposition 1 to obtain a monotone threshold representation for each menu
and let {qxA }x∈A denote the collection of thresholds. We give the sufficiency argument for cross-
menu monotonicity. The argument for the equality conditions, i.e., (i), (ii) in the definition of
(MON), is nearly identical and omitted. Fix A ⊆ B, and let x ∈ A. Note that if x is maximal
(w.r.t. ≤) in A there is nothing to show. Hence, consider xA+ ( = x) ∈ B. Let N qxB denote the min-
imal decisive coalition on the pair {x, b} when the menu is B and b is maximal in B. Standard
arguments (i.e., using strategy-proofness) show that the minimal decisive coalition on {x, xA+ }
+
x,xA
(when the menu is B) also has size N qxB . Let B denote a two-peaked profile supported on B
x,x +
with NqxB voters with peak at x, and let A A denote the profile restricted to A. Dynamic consis-
x,x + x,x +
tency implies that A (A A ) = B (B A ) = x. Hence, N qxA (the minimal decisive coalition)
has size at most NqxB . We now check that property (BD) holds, i.e., that qxA ∈ [qx (A), qx ]. Pro-
ceed by downwards induction on the cardinality of the choice problem A.
Base Step: |A| = |X| − 1.
For the base step consider any menu of cardinality |X| − 1, i.e., where we just delete one
element from X. We check that qxA ≥ qx + . If z < x or z > x + (where z is the element removed
from X to reach A), then qx (A) = qx + and qxA = qx ≥ qx + . Hence, consider z = x + and as-
+
sume, towards contradiction, that qxA < qx (A). Consider a profile x,xA with N qxA voters with
+
peak at x. Note that X (x,xA ) = xA+ since we are alleging that qx + = qx (A) > qxA . OTOH,
+
A (x,xA ) = x – contradicting DC. This verifies the threshold bounds in the base case.
Inductive Step: |A| = |X| − k, k > 1.
Assume the threshold bounds hold for all menus of cardinality |X| − l, l < k. Take x ∈ A

and let A = A \z. We check that qxA ≥ qx (A). Consider xA+ , xA+ . If xA+ = xA
+
, then qxA = qxA .

By the induction hypothesis, qxA ≥ qx (A ). By definition of the numbers qx (·) and cross-menu
monotonicity, we have qx (A ) ≥ qx (A). Hence,

qxA = qxA ≥ qx (A ) ≥ qx (A).
 M. Chandrasekher / Journal of Economic Theory 160 (2015) 175–187 187


Next consider y = xA+ = xA+ = z (so that y > z). We first claim that qxA ≥ qzA . Towards con-

tradiction, say that qxA < qzA . Select a profile x,y with qxA voters with peaks at x and note

that we have A (x,y ) = y. OTOH, A (x,y ) = x, contradicting DC. Hence, qxA ≥ qzA . Now
ad
recall that qx (A) := maxAad qxA+ and let x̂ be an element on which the maximum is attained
and let x̂ ∈ A ∪ {x̃}, where x̃ ∈ X\A. If x̂ = x̃, then x̂ ∈ A. Since x̂ = (x)+
Aad
, where we put
Aad = A ∪ {x̃}, the fact that x̂ ∈ A means that (x)+ = xA+ . Hence, qxA = qxA . Since qxA ≥ qx̂A
ad ad ad
Aad
(as x < x̂), we obtain qxA ≥ qx (A). Now consider x̂ = x̃. If x̂ = z = x̃, then the preceding ar-

gument showed qxA ≥ qzA = qx (A). If x̂ = x̃ = z, then either x < x̃ < z or x̃ > z := xA+ . In the
A∪{x̃} 
former case we find that qxA ≥ qx̃ by arguing as we did to check qxA ≥ qzA .12 Consider the
+
latter case, where z ≤ x̃. Here is where the induction hypothesis enters. Note that we have:

qzA ≥ qz (A ) (by the induction hypothesis)
 ad
:= max qz(A
+
)
(A )ad
A ∪{x̃}
≥ qz +
A ∪{x̃}
≥ qx̃ (since x̃ ≥ z+ )
(since A  A).
A∪{x̃}
≥ qx̃
  A∪{x̃}
Since qxA ≥ qzA , the above string of inequalities gives: qxA ≥ qzA ≥ qx̃ = qx (A). The latter
A∪{x̃}
equality since x̃ was chosen such that qx (A) = qx̃ , i.e., the menu A ∪ {x̃} was the alleged
ad
maximizer of maxAad qxA+ . 2

References

Barberà, S., 2001. An introduction to strategy-proof social choice functions. Soc. Choice Welf. 18, 619–653.
Barberà, S., Massó, J., Neme, A., 1997. Voting under constraints. J. Econ. Theory 76, 298–321.
Bernheim, B.D., Ray, D., 1989. Collective dynamic consistency in repeated games. Games Econ. Behav. 1, 295–326.
Jackson, M.O., Yariv, L., forthcoming. Collective dynamic choice: the necessity of time inconsistency. Am. Econ. J.
Microecon.
Lagunoff, R., 2009. Dynamic stability and reform of political institutions. Games Econ. Behav. 67, 569–583.
Machina, M., 1989. Dynamic consistency and non-expected utility models of choice under uncertainty. J. Econ. Lit. 27,
1622–1668.
Moulin, H., 1980. On strategy-proofness and single peakedness. Public Choice 35, 437–455.
Sen, A., 1969. Quasi-transitivity, rational choice and collective decisions. Rev. Econ. Stud. 36, 381–393.
Siniscalchi, M., 2011. Recursive vector expected utility. Mimeo.

12 Proceeding towards contradiction, say that q A < q A∪{x̃} and consider a profile x,y with q A voters with peak at x.
x x̃ x
A∪{x̃}
If qxA < qx̃ , then A∪{x̃} (x,y ) = y. OTOH, A (x,y ) = x, contradicting DC.