Aspiration-Based Choice Theory∗

Begum Guney† Michael Richter‡ Matan Tsur§

January 22, 2016

Abstract

Numerous studies and experiments suggest that aspirations for desired

but perhaps unavailable alternatives influence decisions. A common

finding is that an unavailable aspiration steers agents to choose simi-

lar available alternatives. We propose and axiomatically characterize a

choice theory consistent with this finding. Similarity is modeled using

a subjective metric derived from choice data. This model offers novel

implications for (1) the effect of past consumption on current decisions,

(2) social influence and conformity within a network, and (3) the dis-

tribution of welfare when firms compete for aspirational agents.

∗We are indebted to Efe A. Ok for his continuous support and guidance throughout thisproject. We especially thank Andrew Caplin, Alan Kirman, Paola Manzini, Marco Mariotti,Vicki Morwitz, David Pearce, Debraj Ray, Ariel Rubinstein, Tomasz Sadzik, Karl Schlag,Andrew Schotter and Roee Teper for their useful suggestions and comments.

†Department of Economics, Ozyegin University. E-mail: begum.guney@ozyegin.edu.tr‡Department of Economics, Yeshiva University. E-mail: michael.richter@yu.edu§Department of Economics, University of Vienna. E-mail matan.tsur@univie.ac.at

1

1 Introduction

There is widespread evidence that decision makers do not always behave as

utility maximizers. While violations of rationality occur, they do not appear to

be random but often follow systematic patterns. To better understand these

patterns, the conditions under which they occur and their economic impli-

cations, a growing literature has proposed choice models that accommodate

departures from rationality. In this paper, we introduce a choice model which

focuses on a different form of bounded rationality: aspirations for unavailable

alternatives.

The question of how unavailable alternatives influence decision making has

received scant attention in the decision theory literature, even though nu-

merous studies and experiments suggest that they have an important effect.

Consider a few examples: (1) The high price (and hence unavailability) of lux-

ury brands leads consumers to purchase cheaper counterfeit products instead.

(2) Individuals may attempt to “keep up with the Joneses”, even though (or

perhaps because) those alternatives are not feasible for them. (3) An employer

may change her ranking of job applicants after interviewing a “superstar” who

is clearly out of reach. (4) Past consumption can create habits that influence

current consumption, especially when past consumption levels are no longer

attainable. (5) Everyday experience suggests that a “daily special” at a restau-

rant that has sold out may impact a consumer’s choice.

Farquhar and Pratkanis (1992, 1993) were the first to experimentally ex-

amine the effect of phantoms – desired but unavailable alternatives – on choice.

Subjects chose between two alternatives T and R as in Figure 1. They found

that adding a phantom P which dominates T but not R, increased the propor-

tion of subjects that chose T (the closest alternative to P ). Moreover, High-

house (1996) and Pettibone and Wedell (2000) found that this aspirational

effect is equal in magnitude to the well-known attraction and compromise

effects.

2

!

!

!

!

!

!

!

!

!

! !

! !

! !

!

! !

!

!

!

!

R!

T!

P!

Attribute!1!

Attribute!2!

Figure 1: Two Alternatives and a Phantom

The above examples and experimental findings present a basic violation

of rationality as irrelevant alternatives affect decisions. However, there are

two systematic patterns: (1) the unavailable alternatives that influence choice

are desired by the decision maker, and (2) when a highly desired alternative

is unavailable, decision makers tend to choose similar alternatives. The first

part of this paper proposes a choice model that captures these features and

provides an axiomatic foundation. The second part studies aspirational agents

in several economic settings.

There are many important economic papers based upon reference-dependent

decision makers. In some of these works, perhaps most notably in Veblen

(1899), Duesenberry (1949), and Pollak (1976), these references are naturally

interpreted as aspirations. There are two main contributions to the present

work. First, to our knowledge, this is the first revealed preference foundation

of how aspirations affect choice. Second, we employ this theory to shed a new

light on three classic economic questions: how does past consumption affect

current decisions, when will social influence in a network lead to conformity

and what are the welfare implications when firms compete for aspirational

buyers?

The choice model extends the standard domain to also include unavailable

alternatives. Formally, a choice problem is a pair of sets (S, Y ) where S ⊆ Y .

3

The set Y consists of the observed alternatives, the subset S is the set of avail-

able alternatives and consequently Y −S is the set of unavailable alternatives.

For example, in Figure 1, Y = {P,R, T}, S = {R, T} and P is the unavailable

alternative. An agent’s choice correspondence assigns a subset of the available

alternatives to each choice problem.

We take an axiomatic approach and characterize a choice procedure defined

by two primitives: a linear order over alternatives and an endogenous metric.

When faced with a choice problem (S, Y ), a decision maker focuses on the

maximal element she observes in Y according to the linear order. We call

this her “aspiration”. She selects the closest available alternative to it where

distances are measured using her metric. Notice that since an alternative is

closest to itself, an aspiration is always selected when available, but when it is

not, it steers the agent towards similar available options.

In the previously discussed experiments, the alternatives were points in

a Euclidean space and there is a natural notion of similarity (the Euclidean

metric). However, measuring similarity in real life may not be so straightfor-

ward and these assessments are often subjective – two diamonds may appear

indistinguishable to one decision maker, but couldn’t be further apart to an-

other. Therefore, the model derives the metric endogenously from choice data,

as standard utility functions are. This introduces a technical hurdle because

a metric must satisfy additional properties (reflexivity, symmetry and the tri-

angle inequality) that a utility function need not. Remarkably, its characteri-

zation relies on three simple and intuitive axioms.

We study the economic implications of this model in three environments.

The first is a standard consumption setting where the agent remembers her

previously consumed bundles and therefore may aspire to them. In this setting,

following a decrease in income, previously consumed bundles are no longer

affordable but will have an impact as aspirations. On the other hand, following

an increase in income, previously consumed bundles remain affordable and

therefore will not affect current consumption. We are not aware of any habit-

based consumption models that make this intuitive distinction.

The second is a network setting where agents make choices and observe

4

the choices of their neighbors. The model provides a novel channel of social

influence: one agent’s choice may serve as the aspiration for another agent,

whose choice in turn affects a third agent and so on throughout the network.

A simple condition guarantees the existence of a unique steady state. We also

examine the likelihood of conformity , the event that agents facing different

options make similar choices. If agents are rational, then the likelihood of

conformity vanishes as the number of agents increases. However, if agents are

aspirational, this likelihood is uniformly bounded away from zero.

The third environment is a competitive market with profit-maximizing

firms and aspirational buyers. A “red” firm and a “blue” firm both sell

a high- and low-quality good to a continuum of wealth-constrained buyers.

Firms engage in price competition. Buyers that cannot afford the high-quality

goods will aspire to the choices of richer buyers who can. Thus, firms have

an added incentive to attract these richer buyers in order to create aspirations

for its goods. The unique equilibrium has the following features: (1) for a

good to serve as an aspiration, it need not only be better, but the quality

gap must be sufficiently large; (2) firms lose money on their aspirational good

and cross-subsidize these losses with profits from the low-quality market; and

(3) suprisingly, welfare can either increase or decrease, but the distribution of

welfare always changes in favor of the richer buyers.

The rest of the paper is organized as follows: Section 2 discusses the related

theoretical literature. Section 3 presents the model and the main representa-

tion result along with a discussion of the experimental literature. Section 4

studies the three economic environments and Section 5 concludes. Two ap-

pendices contain proofs and extensions of the main theorem.

2 Related Literature

The notion that an aspiration towards some higher goal influences behavior

is widespread across the social sciences (see Siegel (1957), Appadurai (2004)

and Ray (2006)). Examples range from studies concerning occupational choice

(Correll 2004) and educational attainment (Kao and Tienda 1998) to general

5

happiness (Stutzer 2004) and risk-taking behavior (Ray and Robson 2012).

In various economic models, utility functions depend upon a reference

point, which may be interpreted as an aspiration or an aspirational thresh-

old. In a recent paper, Genicot and Ray (2015) study growth and income

inequality in an overlapping generations model where an aspiration for the

next generation’s wealth level influences today’s consumption-savings trade-

off. Borgers and Sarin (2000) consider a reinforcement learning model where

an agent aspires to a payoff threshold which endogenously changes with re-

alized payoffs and the probability with which an action is repeated depends

upon whether this benchmark is met or not. Karandikar, Mookherjee, Ray,

and Vega-Redondo (1998) study similar behavior in a repeated game setting.

A contribution of the present work is providing a revealed preference theory

of aspirations. Furthermore, the model is general and portable to a wide range

of settings, including those with non-monetary outcomes. In contrast, previous

work focused on specific settings and ad-hoc definitions of aspirations. The

model developed here may prove useful in relating disparate observations from

a variety of settings.

The most closely related papers in decision theory are Masatlioglu and Ok

(2005, 2014) and Ok, Ortoleva, and Riella (2015). The former takes a choice

problem to be a set of available alternatives plus a status quo alternative

and the latter endogenizes the selection of the status quo. In all of these

papers, the status quo alternative is available for choice, but may rule out other

alternatives from consideration. The aspirational effect differs with respect to

both of these features: aspirations are often unavailable and influence choices

via similarity.

The framework that we develop consists of choice problems with frames,

as in Rubinstein and Salant (2008) and Bernheim and Rangel (2009). The

procedure that we characterize can be thought of as taking place in two

stages, and thus tangentially relates to other two-stage choice procedures,

such as: triggered rationality (Rubinstein and Salant 2006), sequential ra-

tionality (Manzini and Mariotti 2007), limited attention (Masatlioglu, Naka-

jima, and Ozbay 2012) and the warm glow effect (Cherepanov, Feddersen, and

6

Sandroni 2013). The current model differs significantly from these models with

respect to both framework and procedure.

Outside of decision theory, Rubinstein and Zhou (1999) axiomatize a bar-

gaining solution which selects from a utility possibility set the point closest to

an external reference. This bargaining solution resembles our choice procedure

except that the metric and reference point are exogenous in their model.

In general, there are several models where distance to a reference point de-

termines decisions. The canonical example is prospect theory (Kahneman and

Tversky 1979) and a more recent one is salience theory (Bordalo, Gennaioli

and Shleifer, 2012a, 2012b), where the salience of an attribute is determined by

its distance from the average observed level of that attribute. These theories

require clear measures of distance and therefore are typically applied to mon-

etary outcomes, probabilties, or vectors in RN . The current model may prove

useful for extending distance-based theories to settings without such clear-cut

distances.

3 The Choice Model

We begin with some preliminary definitions. A compact metric space X

(finite or infinite) is the grand set of alternatives. Let X denote the set of

all nonempty closed subsets of X. X is endowed with the Hausdorff metric

and convergence with respect to this metric is denoted byH→. A linear order

is a reflexive, complete, transitive and anti-symmetric binary relation.

A choice problem is a pair of sets (S, Y ) where S, Y ∈ X and S ⊆ Y . The

set of all choice problems is denoted by C(X) and a choice correspondence is

a map C : C(X) → X such that C(S, Y ) ⊆ S for all (S, Y ) ∈ C(X). When

an agent faces a choice problem (S, Y ), she observes the potential set Y , but

chooses from the feasible set S. We follow previous literature and use the term

phantom for an observed but unavailable alternative, that is, an alternative

in Y but not in S. An unobserved alternative is not choosable and hence the

requirement that S ⊆ Y .

Note that unavailable alternatives constitute part of the choice data. That

7

is, they are observable not only to the decision maker, but also to the economist.

This is plausible in a variety of settings. For example, consider a high school

student applying to colleges. Both the schools she has applied to (the

potential set) and the schools she has been admitted to (the feasible set) are

externally observable. The same is true for a job applicant applying through

a career center, customers shopping online and the examples mentioned in the

introduction.

We take an axiomatic approach. When unavailable alternatives impact

choices, the Weak Axiom of Revealed Preference (WARP) is necessarily

violated. The first two axioms restrict the domains in which such violations can

occur: the first requires that agents behave rationally when all the observed

alternatives are also available and the second assumes rational behavior when

a potential set is held fixed.

Axiom 3.1 (WARP for Phantom-Free Problems) For any S, Y ∈ X such that

S ⊆ Y ,

C(S, S) = C(Y, Y ) ∩ S, provided that C(Y, Y ) ∩ S 6= ∅.

Axiom 3.2 (WARP Given Potential Set) For any (S, Y ), (T, Y ) ∈ C(X) such

that T ⊆ S,

C(T, Y ) = C(S, Y ) ∩ T, provided that C(S, Y ) ∩ T 6= ∅.

Thus, violations of rationality are due to unavailable alternatives and can

only occur when potential sets vary. These restrictions allow for the previ-

ously discussed experimental evidence that choice reversals may occur when

phantoms are introduced.

We interpret aspirations as the choices that would be made in the absence

of feasibility restrictions, that is, when the feasible and potential sets coincide.

Under this interpretation, the following axiom states that if two potential sets

generate the same aspirations, then they will influence choices in the same

way.

8

Axiom 3.3 (Independence of Non-Aspirational Alternatives)

For any (S,Z),(S, Y ) ∈ C(X),

C(S,Z) = C(S, Y ), provided that C(Z,Z) = C(Y, Y ).

A choice correspondence is called aspirational if it satisfies Axioms 3.1-3.3.

The next two axioms are technical. The first guarantees a unique aspiration

in any given problem. The second is a standard continuity axiom, which

ensures the existence of maximums and trivially holds when the grand set is

finite. Appendix B considers relaxations of these two axioms.

Axiom 3.4 (Single Aspiration Point) |C(Y, Y )| = 1 for all Y ∈ X .

Axiom 3.5 (Continuity) For any (Sn, Yn), (S, Y ) ∈ C(X), n = 1, 2, . . . and

xn ∈ Sn for all n such that SnH→ S, Yn

H→ Y and xn → x,

if xn ∈ C(Sn, Yn) for every n, then x ∈ C(S, Y ).

The above axioms yield our main representation result:

Theorem 3.1 C satisfies Axioms 3.1-3.5 if and only if there exists

1. a continuous linear order � and

2. a continuous metric d : X2 → R+

such that

C(S, Y ) = argmins∈S

d(s, a(Y )) for all (S, Y ) ∈ C(X),

where a(Y ) is the �-maximum element of Y .

In the above procedure, from the choice problem (S, Y ) a decision maker

chooses the closest available alternative to her aspiration a(Y ). Her aspiration

is the maximum element (according to the linear order �) that she observes

9

and closeness is measured using her metric d. This metric is derived endoge-

nously and need not coincide with the metric that the grand set is originally

endowed with. Generally, agents making different choices will be represented

with different metrics. Notice that if an aspiration is available a(Y ) ∈ S, then

it is uniquely chosen C(S, Y ) = a(Y ), because an alternative is always closest

to itself.

Remark 1: In the standard choice model, only feasible alternatives are

relevant for choice and thus constitute all the necessary choice data. Here,

we also include unavailable alternatives, which are externally observable in

many settings as previously discussed (e.g. college/job applications, online

shopping, past consumption). Notice that aspirations, unlike other reference

points (such as status quos), are typically unobservable and we therefore do

not directly include them as choice data.

Remark 2: In this framework, for an alternative to be choosable it must

be observable, and therefore S ⊆ Y . An equivalent formulation would drop

the subset restriction and take aspirations to be drawn from S ∪ Y . In this

formulation, if S and Y do not intersect, then Y is the set of unavailable

alternatives. Subject to relabeling, the main representation theorem holds as

is.

Remark 3: The axioms are also consistent with a reference-dependent utility

maximization procedure, where C(S, Y ) = maxs∈S

U(s,max(Y,�)). The repre-

sentation used here offers several advantages: First, if |X| = n, then these

utility functions have n2 degrees of freedom whereas a distance function has

less than half that, i.e. (n2 − n)/2. Thus, the distance minimization pro-

cedure is better identified than the reference-dependent utility maximization

procedure. Second, distance minimization is consistent with the experimental

findings discussed in the next subsection. Finally, distance minimization offers

a geometric interpretation that is natural in many settings, as illustrated in

Section 4.

Remark 4: A rational agent who maximizes a single continuous utility func-

tion over all available alternatives satisfies the above axioms and therefore, the

10

rational model is nested. Representing the choices of such an agent through

distance minimization is trivially possible as utility distances, d(a, b) = |U(a)−U(b)|. However, when agents are irrational, potential sets influence choices and

such a trivial representation does not exist.

Remark 5: Nishimura and Ok (2014) show that the existence of a continuous

choice function on a compact metric space implies that it is of topological

dimension 1. Thus, the existence of an aspiration function imposes restrictions

on the space. This dimensionality restriction has no impact when X is a

finite set, but has consequences when X is a larger path-connected space.

In Appendix B, we consider relaxations of both the Continuity and Single

Aspiration Point Axioms where this dimensionality restriction does not apply.

3.1 Experimental Evidence

Various experiments study the impact of desired unavailable alternatives, “phan-

toms”, on choice. In a typical experiment, subjects choose between two al-

ternatives {T,R} and treatments vary both the presence and location of a

phantom alternative P (see Figure 1). Farquhar and Pratkanis (1992, 1993)

find that the addition of a phantom P , significantly increases the fraction of

subjects who choose T (the closer alternative to P ). Guney and Richter (2015)

find additional evidence for this aspirational effect.

Highhouse (1996) compares this aspiration effect to the well-known

attraction effect that introducing a decoy D which is dominated by exactly one

alternative increases the choice share of that dominating alternative (Huber,

Payne, and Puto 1982). His experiment varies the alternative targeted (T or R

in Figure 2) and the targeting effect: in one treatment, there is a a dominating

phantom P (the aspiration effect), in the other, there is a dominated decoy

D (the attraction effect). He finds that targeting an alternative (with either

a decoy or a phantom) increases the frequency with which it is chosen and no

statistically significant difference in the magnitude of the two effects.

Pettibone and Wedell (2000) similarly compare the aspiration effect to

the compromise effect. In their treatments, the unavailable alternative either

11

!

!

!

!

!

!

!!!!

!

!

! !

! !

! !

!

! !

!

!

!

!

R!

T!

P!

!!Attribute!1!

Attribute!2!

!!!!D!

Figure 2: Two Choice Alternatives with a Phantom or a Decoy

dominates the target alternative or is placed so that the target becomes a

compromise alternative. Like Highhouse (1996), they find that both treat-

ments significantly increase the proportion of agents who choose the target T

and there is no statistically significant difference in the magnitude of the two

effects.

These findings may seem surprising to readers familiar with the

attraction effect. An intuitive explanation for the attraction effect is that

a dominance relation makes the dominating alternative more attractive and

hence the dominated alternative less. This intuition implies that a dominat-

ing phantom would make the target less attractive and therefore less subjects

would choose it. But, the opposite happens in the above experiments, which

suggests that different mechanisms may underly the aspiration and the attrac-

tion effect. Theoretically, there are several ways to extend our choice model

to incorporate both, but in order to do so, further evidence would be useful in

understanding the interaction of these two effects.

4 Economic Implications

This section studies three environments. The first is a standard consumption

setting where agents may aspire to past consumption bundles. The second is a

network model where agents may draw their aspirations from their neighbors’

12

choices. The third is a market where firms compete for aspirational buyers.

4.1 Aspirations for Past Consumption

We start with a standard consumption problem: an agent chooses a pair (x, y)

where x is a quantity of a nondurable good and y is the expenditure on all

other goods. The budget set today is B1 = {(x, y) : px + y = I} where I

is income and p is the price of the good and the set B0 denotes previously

consumed bundles. The agent’s feasible set is her budget set S = B1, while

her potential set also includes alternatives consumed in the past Y = B0∪B1.

The agent has continuous, strictly monotonic and strictly convex preferences

and measures distances using the Euclidean metric.

The main idea is depicted in Figure 3. As a benchmark, take a rational

agent who faces the same budget set B1 and has the same underlying prefer-

ence. This agent will choose point c, while the choice of the aspiration agent

depends on where her aspiration point a = max(B0 ∪B1,�) is located:

• If a is in region 1, she will choose a point on the budget line below c.

• If a is in region 2, she will choose a point on the budget line above c.

• Otherwise, she will choose c.

Good x

Good y

c

1

2

Good x

Good y

c

1

2

c’

Figure 3: The Persistent Effect of Past Consumption

13

This observation has several interesting implications regarding how agents

respond to income and price shocks. For example, when income increases,

B0 ⊂ B1, previous bundles are still affordable and therefore will not affect

decisions. In contrast, when income decreases, past consumption bundles are

no longer affordable and will affect decisions. Thus, an aspirational agent

responds rationally to positive income shocks but may respond irrationally to

negative income shocks.

To determine the direction of the aspirational effect, consider an agent

with demand x(p, I), y(p, I) who undergoes an infinitesimal income or price

shock. If income increases or price decreases, then the behavior of the rational

agent and aspirational agent coincide. Otherwise, the aspirational agent will

consume more x than her rational counterpart precisely when:

If incomes decreases:

(I

p,−I

)·(∂x

∂I,∂y

∂I

)> 0

If price increases:

(I

p,−I

)·(∂x

∂p,∂y

∂p

)< 0

For another implication, we analyze the post-sale demand for a good. A

number of empirical studies document the difference between pre-sale and

post-sale consumption levels of a good.1 During the sale, the agent’s budget

set expands and her choice will change. After the sale ends, the budget set

returns to what it was and an aspirational agent’s choice will depend upon her

consumption during the sale (see Figure 3, right panel).

Proposition 4.1 There is always a sale price p′ < p such that a temporary

sale will increase the post-sale consumption level of the good.

The aspirational effect offers a novel explanation for the pre- and post-sale

consumption difference and shows that a low enough sale price will have a

positive long-term effect.

1See, for example, Van Heerde, Leeflang, and Wittink (2000), Hendel and Nevo (2003)and DelVecchio, Henard, and Freling (2006).

14

Remark: The analysis here uses the Euclidean metric to illustrate the

application, but if the agent were to employ a different metric, then the shape

of the regions in Figure 3 could be suitably altered.

Related Literature: In habit formation models, an agent’s choices may

depend upon her past consumption in various forms (for example, see Pol-

lak (1970)), but the aspiration effect has a distinctive one-sided feature: only

superior previously consumed bundles may affect choices. Recently, Gilboa,

Postlewaite, and Samuelson (2015) study a decision-theoretic model of mem-

orable goods. Agents are forward-looking and solve a dynamic optimization

problem where extraordinary consumption levels of memorable goods have

long-lasting utility effects. As a result, such consumption levels are typically

not immediately repeated. In contrast, an aspirational agent would seek sim-

ilar extraordinary outcomes. In this sense, these models are readily distin-

guishable in their consumption patterns: aspirational agents exhibit addiction

and memorable agents may consume memorable goods periodically.

4.2 Aspirations in Networks

There are N agents connected in a network Γ. Agents i and j are neighbors if

(i, j) ∈ Γ. Each agent faces a feasible set Si and these sets are non-intersecting.

An agent observes the choices of her neighbors. Therefore, if each agent j

chooses xj, then agent i faces the potential set Yi = Si ∪ {xj : (i, j) ∈ Γ}and her choice problem is (Si, Yi). We assume that all agents have a common

aspirational ranking and the distance functions are such that no two goods are

equidistant from a third good, i.e. for all i and goods x, y, z: di(x, y) 6= di(x, z).

In this model, one agent’s choice may serve as the aspiration of another

agent, whose choice in turn affects a third agent and so on throughout the net-

work. An allocation (x1, . . . , xN) is a steady state if no agent would change his

choice given the choices of all other agents, that is, ∀i,xi = C(Si, Yi).

Proposition 4.2 A steady state exists and is unique.

15

The idea of the proof is as follows. Since all agents share the same aspi-

rational ranking, we can order the agents according to their most preferred

feasible alternative. Notice that an agent never aspires to the choices of those

lower-ranked than him. Thus, if agents choose in sequence according to this

order, the final outcome will be a steady state. Uniqueness is proved similarly.

The following example demonstrates that if agents have different aspiration

rankings then equilibrium may not exist.

Example: Consider two agents with feasible sets S1 = {a, b}, S2 = {x, y}and identical distance functions where d(x, b) < d(b, y) < d(a, y). However,

their aspiration rankings differ: y �1 a �1 b �1 x and b �2 y �2 x �2 a.

The allocation (a, y) where each agent chooses her highest-ranked available

alternative is not a steady state because agent 1 aspires to y and thus would

choose b instead of a. But, (b, y) is also not a steady state because agent 2

aspires to b and switches to x. Since agent 1 does not aspire to x and ranks a

above b, (b, x) is not a steady state either. In short, the cycle (a, y)1→ (b, y)

2→(b, x)

1→ (a, x)2→ (a, y) rules out the existence of a steady state.

We now demonstrate some properties of the steady state. Consider a world

with many goods each of which is either red or blue, but they may differ in

several other dimensions (such as price, quality, or size). The significance of a

good’s color is that any two goods of the same color are closer to each other

than to a third good of the other color. For simplicity, we let each choice set

contains two goods, one of each color.

Definition: A steady state exhibits conformity if all agents choose a good of

the same color.

The agent with the top-ranked alternative (amongst all available alterna-

tives) will always choose it and her neighbors will be influenced to choose that

color. Thus, in the complete network, conformity occurs. On the other hand,

in the empty network (when no agents are linked), each agent chooses inde-

pendently and conformity is highly unlikely. All other networks are between

these two extremes and we analyze the average behavior across them. Thus,we

hold the choice sets fixed and take any network and linear aspiration order to

16

be equally likely and drawn independently. We find that, on average, there is

a non-trivial probability of conformity.

Proposition 4.3 The probability of conformity is uniformly bounded away

from 0 for any number of agents.

The logic of the proof is as follows. In any steady state, some agents may be

influenced by others. Since the color chosen by influenced agents is determined

by other agents, only the color chosen by uninfluenced agents is random. The

expected number of agents that are uninfluenced is monotonically increasing

in N , but we establish a finite upper bound. Therefore, on average, choices

are random for a finite number of agents and there is a non-trivial probability

of conformity.

The above proposition demonstrates a failure of the strong law in our

setting as extreme outcomes do not vanish. In contrast to aspirational agents,

rational agents choose independently and conformity occurs with a vanishing

probability, 1/2N−1.

Aside from the question of conformity, the model provides interesting in-

sights into questions of social influence,2 trend-setters and wealth. As in most

network models, an agent’s social influence depends upon her location, but

in this case, it also depends upon the quality of the goods which that agent

visibly consumes. An immediate implication is that wealthier agents will have

more social influence because they can afford better options. The application

considered in the next subsection involves sophisticated firms who strategically

target these richer agents in order to increase demand for their products.

Related Literature: The key feature that distinguishes this network model

is that an agent can be directly influenced by at most one agent, specifically,

the one who chooses the maximal observed alternative. Which agent turns

out to be the influencer depends endogenously upon both the network and

preferences. In contrast, in many network models agents’ payoffs/choices de-

pend upon all of their neighbors’ choices. This is typically the case in network

2A measure of agent i’s social influence is the number of agents that change their choiceswhen agent i is removed from the network.

17

games, word-of-mouth communication and other social learning models (see

Bala and Goyal (2000), Jackson and Yariv (2007), Jackson (2008), and Camp-

bell (2013)).

Frequently, network models are plagued with a multiplicity of equilibria.

For example, if agents imitate the median action of their neighbors, which in

our two-color setting means that agents choose the most popular color that

they observe, then multiple equilibria arise: a “red” one, a “blue” one and

potentially many mixed ones as well. In contrast, the directed influence of our

model always delivers a unique steady state.

Conformity may occur for many reasons. Agents may imitate the actions

of those who are better informed (Banerjee 1992, Bikhchandani, Hirshleifer,

and Welch 1992), there may be positive strategic externalities (Morris 2000),

or agents may conform to a public signal (Bernheim 1994, Pesendorfer 1995).

Perhaps more closely related are models of disease contagion where agents can

be either healthy or sick and sick agents may infect healthy neighbors. Thus,

the infection of one agent by another parallels the influence of one agent on

another and the probability of a pandemic corresponds to that of conformity.

There are however two major differences: in the aspirational model, infection

travels in only one direction which is determined by whoever has the superior

alternative and there may be competing sources of infection, namely the “red”

and “blue” infections.

4.3 Competition over aspirational buyers

In this subsection, we study a competitive market with profit-maximizing firms

and aspirational buyers. Each firm sells goods of two different quality levels

qH > qL with costs 1 > cH > cL > 0 and qL − cL > 0. Some firms can

differentiate themselves by brand and others cannot. There are at least two

firms of each type. For expositional clarity, we consider two branded firms

“red” and “blue” and refer to the generic firms as “colorless”. Firms engage

in price competition.

There is a continuum of buyers with wealth w distributed uniformly on

18

[0, 1], each of whom cannot pay more than their wealth. Buyers are aspirational

and each observes the choices of all others (a complete network). Therefore,

as in the previous subsection, a buyer’s feasible set consists of the goods that

she can afford and her potential set also includes the goods chosen by others.

The aspirational ranking is determined by u = q − p and each buyer breaks

ties randomly. Similarity is lexicographic first by color and then by utility

(that is, two goods of the same color are always closer to each other than to a

third good of a different color). A buyer who aspires to a red good will always

purchase a red good if one is affordable and if not, will purchase a maximal

utility good amongst those that are affordable.

As a benchmark, consider rational buyers with the utility function u = q−p.In equilibrium, Bertrand-competition drives prices down to marginal costs and

firms break even. If the high-quality good is more efficient, qH − cH > qL− cL,

then both goods are sold at marginal costs, cH and cL. If the low-quality good

is more efficient, then only this good is sold at cL.

However, when buyers are aspirational, both goods being sold at marginal

cost is not an equilibrium. The reason is that buyers who cannot afford the

high-quality good will aspire to it, and therefore the red firm can profitably

increase the price of its low-quality good without losing all of these buyers.

In fact, the firm can do even better by slightly decreasing the price of its

high-quality good to capture the entire high-quality market and consequently

monopolize the low-quality market.

Proposition 4.4 When buyers are aspirational, there is generically a unique

pure strategy NE.

1. If qH − qL > g(cH , cL), the red and blue firms charge the same prices

p∗L, p∗H where cL < p∗L < p∗H < cH , the uncolored firms sell only the

low-quality good at price cL, and all firms break even.

2. If qH − qL < g(cH , cL), then only the low-quality good is sold and the

price is cL.

and 0 < g(cH , cL) < cH − cL.

19

To understand the proposition, consider two cases. First, when the high-

quality good is more efficient, then case 1 holds because qH − qL > cH − cL >g(cH , cL). In equilibrium, both goods are sold and prices are jointly determined

by:

(i) Monopoly pricing of the low-quality good pL = (pH + cL)/2.

(ii) Zero-profit condition (1− pH)(pH − cH) + (pH − pL)(pL − cL) = 0.

The prices p∗L, p∗H are the unique admissible solution to the above equations.

The driving force here is the demand externality of the richer buyers on the

poorer ones: the red and blue firms compete for aspirations by lowering the

price of their high-quality good in order to monopolize the low-quality market.

This process settles when the losses in the high-quality market are exactly

offset by the monopoly profits in the low-quality market. Buyers with wealth

above p∗H purchase the red or blue high-quality good. Buyers with lower wealth

cannot afford these high-quality goods, but aspire to them. Buyers in region

II purchase either the red or blue low-quality goods at the price p∗L and buyers

in region I purchase a colorless low-quality good at the price cL.

IIIIIIIV!! 0 1 !! !!∗ !!∗

Figure 4: Competition for aspirational buyers

Second, when the low-quality good is more efficient, whether both goods

are sold depends upon the quality gap qH − qL. If only the low-quality good

is sold, it must be sold at cost. Notice that the red firm may still introduce a

high-quality good to monopolize the low-quality market. To create aspirations,

the red firm must price its high-quality good low enough to attract some buyers

(incurring losses), and this deviation is profitable when there are sufficiently

many buyers in the low-quality market. When the quality gap is small (case

2), any price which attracts buyers to the high-quality market will cannibalize

the low-quality market and this deviation is not profitable. But, when the gap

is large (case 1), this is a profitable deviation.

20

We now compare the social welfare of the aspirational and rational equi-

libria. Notice that they differ only when the quality gap is large. In the aspi-

rational equilibrium, poorer buyers who purchase the red or blue low-quality

goods cross-subsidize richer buyers who purchase the red or blue high-quality

goods. The utility gains of the richer buyers are larger than the utility losses

of the poorer buyers exactly when the high-quality good is more efficient.

Proposition 4.5 Social welfare is higher in the aspirational equilibrium than

in the rational equilibrium if and only if qH − cH > qL − cL.

In both the rational and aspirational models firms break even and therefore

a welfare comparison can focus solely on the change in allocations. When the

high-quality good is more efficient, allocations change only for buyers with

wealth in region III (Figure 4). These buyers purchase the more-efficient high-

quality good instead of the less-efficient low-quality one and welfare increases.

Conversely, when the high-quality good is less-efficient, since buyers in regions

III and IV switch their purchases to it, welfare decreases.

Thus, overall welfare may increase or decrease, but the welfare distribution

always changes to the benefit of richer buyers. In general, wealthier buyers

are better off because they have more options, we find that competition for

aspirations amplifies this advantage, increasing the welfare gap between rich

and poor.

There are a number of anecdotal examples of “premium loss leaders”, high-

quality goods sold beneath costs in order to increase demand for other prod-

ucts. Examples include: high-end low-volume cars produced alongside more

affordable and similar options,3 designer dresses given to celebrities which pos-

itively impact their mass market offerings, and restaurant empires anchored by

a money-losing highly-rated flagship which is subsidized by profitable cheaper

franchises. An article in Fortune4 explains:

“For many Michelin star restaurateurs, the restaurant is a loss

leader whose fame allows the chef to charge high speaking or private

3In 2000, 1,371 BMW Z8s were sold while 511,052 similar 3-series cars were sold.4“The Curse of the Michelin-Star Restaurant Rating”, Fortune, December 11, 2014.

21

cooking fees; others start lines of premade food or lower priced

restaurants. In a recent interview, David Munoz of DiverXo told

me that 2015 would be the first breakeven year in his company’s

eight-year existence, and then only because of the expansion of his

street-food chain, StreetXo.”

To sum up the main points, we find that there is a unique equilibrium with

the following features: (1) for a good to serve as an aspiration, it need not only

be better, but the quality gap must be sufficiently large; (2) firms lose money

on their aspirational good and cross-subsidize these losses with profits from

the low-quality market; and (3) whether competition for aspirations increases

or decreases welfare depends upon the relative efficiency of the goods, but the

distribution of welfare always changes in favor of the richer buyers.

Related Literature: In markets with sophisticated firms and boundedly ra-

tional consumers, it is natural to expect that manipulation of consumers leads

to welfare losses. While this intuition is indeed confirmed by several models

in the literature(Gabaix and Laibson 2006, Spiegler 2006, Spiegler 2011), the

above demonstrates that this need not always be the case.

This application relates to a few recent papers. In Kamenica (2008), buy-

ers’ preferences depend upon a global taste parameter which the firm knows

but buyers may not. The firm offers a menu of goods and the buyers make

inferences about the parameter through the firm’s menu choice. In equilib-

rium, firms may introduce unprofitable loss leaders to influence the beliefs of

uninformed buyers. In another recent paper, Kircher and Postlewaite (2008)

consider an infinitely repeated search model where firms vary in unobservable

quality and prices are fixed. Since wealthy buyers consume more frequently,

they are better informed, and as a result, the poor have an incentive to imitate

the rich. In equilibrium, firms offer higher quality to wealthier buyers in order

to attract other buyers. In both models, firms may produce superior goods for

the sake of influencing less informed buyers. In our model, the high-quality

good is subsidized for a different reason, to create aspirations.

22

5 Conclusion

A novel choice setting that allows agents’ choices to vary with observed but

unavailable alternatives was introduced. In this setting, we characterize a

distance-minimizing choice procedure that captures a basic aspirational effect

observed in real-life examples and experiments: desired but unavailable alter-

natives (which we call aspirations) influence agents to choose similar available

alternatives. The use of similarity, expressed by an endogenously derived met-

ric, is a key innovation of this paper.

Two important features of any choice theory are its axiomatic foundation

and the economic implications that are deduced. This work place equal weight

on both.

We provided three simple conditions that characterize the aspirational

choice procedure: (1) the agent behaves rationally when every alternative

observed is available, (2) the agent behaves rationally when the alternatives

she observes do not change, and (3) an agent who draws the same aspiration

from two different problems ranks alternatives equivalently. These conditions

are in line with experimental findings and can be easily tested and falsified.

While there are various sources for aspirations, we have focused on two nat-

ural ones: previously consumed bundles and neighbors’ consumption. In each

case, the aspirational model provided novel answers to fundamental questions:

how agents respond to income and price shocks; when and why conformity

of choices occurs; the welfare gains and loses of competition over aspirational

buyers.

There are several questions that may be of interest for future research. The

current framework takes choice problems to be pairs of sets, which enables a

distinction to be made between available and unavailable alternatives. Be-

yond the particular aspirational effect studied here, this framework may also

prove useful in studying other context effects. One possibility is an axiomatic

theory that incorporates both the attraction and aspiration effects within a

single model. In this vein, an experiment that investigates both effects simul-

taneously, rather than separately as in Highhouse (1996), may shed light on

23

both of them. Another interesting avenue is to combine the latter two ap-

plications and study competition over aspirational buyers connected through

more complicated networks. Finally, it may be worth analyzing a dynamic

consumption-savings problem where aspirations are drawn from all past con-

sumption bundles.

24

References

Appadurai, A. (2004): “The Capacity to Aspire: Culture and the Terms of

Recognition,” in Culture and Public Action: A Cross-Disciplinary Dialogue

on Development Policy, ed. by V. Rao, and M. Walton. Stanford University

Press.

Bala, V., and S. Goyal (2000): “A noncooperative model of network for-

mation,” Econometrica, 68(5), 1181–1229.

Banerjee, A. V. (1992): “A simple model of herd behavior,” The Quarterly

Journal of Economics, pp. 797–817.

Ben Yaacov, I., A. Berenstein, and S. Ferri (2011): “Reflexive rep-

resentability and stable metrics,” Mathematische Zeitschrift, 267(1-2), 129–

138.

Bernheim, B. D. (1994): “A theory of conformity,” Journal of political Econ-

omy, pp. 841–877.

Bernheim, B. D., and A. Rangel (2009): “Beyond Revealed Prefer-

ence: Choice-Theoretic Foundations for Behavioral Welfare Economics,”

The Quarterly Journal of Economics, 124, 51–104.

Bikhchandani, S., D. Hirshleifer, and I. Welch (1992): “A theory

of fads, fashion, custom, and cultural change as informational cascades,”

Journal of political Economy, pp. 992–1026.

Bordalo, P., N. Gennaioli, and A. Shleifer (2012a): “Salience and

Experimental Tests of the Endowment Effect,” American Economic Review

Papers and Proceedings, 102, 47–52.

(2012b): “Salience Theory of Choice Under Risk,” The Quarterly

Journal of Economics, 127, 1243–1285.

Borgers, T., and R. Sarin (2000): “Naive Reinforcement Learning with

Endogenous Aspirations,” International Economic Review, 41(4), 921–50.

25

Campbell, A. (2013): “Word-of-mouth communication and percolation in

social networks,” The American Economic Review, 103(6), 2466–2498.

Cherepanov, V., T. Feddersen, and A. Sandroni (2013): “Revealed

Preferences and Aspirations in Warm Glow Theory,” Economic Theory, 54,

501–535.

Correll, S. J. (2004): “Constraints into Preferences: Gender, Status, and

Emerging Career Aspirations,” American Sociological Review, 69, 93–113.

DelVecchio, D., D. H. Henard, and T. H. Freling (2006): “The Ef-

fect of Sales Promotion on Post-Promotion Brand Preference: A Meta-

Analysis,” Journal of Retailing, 82, 203–213.

Duesenberry, J. S. (1949): Income, Saving, and the Theory of Consumer

Behavior.

Evren, O., and E. A. Ok (2011): “On the multi-utility representation of

preference relations,” Journal of Mathematical Economics, 47(4), 554–563.

Farquhar, P. H., and A. R. Pratkanis (1992): “A Brief History of Re-

search on Phantom Alternatives: Evidence for Seven Empirical Generaliza-

tions About Phantoms,” Basic and Applied Social Psychology, 13, 103–122.

(1993): “Decision Structuring with Phantom Alternatives,” Manage-

ment Science, 39, 1214–1226.

Gabaix, X., and D. Laibson (2006): “Shrouded Attributes, Consumer

Myopia, and Information Suppression in Competitive Markets,” Quarterly

Journal of Economics, 121(2).

Genicot, G., and D. Ray (2015): “Aspirations and Inequality,” Working

Paper.

Gilboa, I., A. Postlewaite, and L. Samuelson (2015): “Memory Util-

ity,” Working Paper.

26

Guney, B., and M. Richter (2015): “An Experiment on Aspiration-Based

Choice,” Journal of Economic Behavior & Organization, 119, 512 – 526.

Hendel, I., and A. Nevo (2003): “The Post-Promotion Dip Puzzle: What

Do the Data Have to Say?,” Quantitative Marketing and Economics, 1,

409–424.

Highhouse, S. (1996): “Context-Dependent Selection: The Effects of De-

coy and Phantom Job Candidates,” Organizational Behavior and Human

Decision Processes, 65, 68 – 76.

Huber, J., J. W. Payne, and C. Puto (1982): “Adding Asymmetrically

Dominated Alternatives: Violations of Regularity and the Similarity Hy-

pothesis,” Journal of Consumer Research: An Interdisciplinary Quarterly,

9, 90–98.

Jackson, M. O. (2008): Social and Economic Networks, vol. 3. Princeton

University Press Princeton.

Jackson, M. O., and L. Yariv (2007): “Diffusion of Behavior and Equi-

librium Properties in Network Games,” The American Economic Review,

97(2), 92–98.

Kahneman, D., and A. Tversky (1979): “Prospect theory: An analysis of

decision under risk,” Econometrica, pp. 263–291.

Kamenica, E. (2008): “Contextual Inference in Markets: On the Informa-

tional Content of Product Lines,” The American Economic Review, 98(5),

2127–2149.

Kao, G., and M. Tienda (1998): “Educational aspirations of minority

youth,” American Journal of Education, pp. 349–384.

Karandikar, R., D. Mookherjee, D. Ray, and F. Vega-Redondo

(1998): “Evolving aspirations and cooperation,” journal of economic theory,

80(2), 292–331.

27

Kircher, P., and A. Postlewaite (2008): “Strategic firms and endoge-

nous consumer emulation,” The Quarterly Journal of Economics, 123, 621–

661.

Levin, V. (1983): “Measurable Utility Theorems for Closed and Lexico-

graphic Preference Relations,” Soviet Math. Dokl., 27, 639–643.

Manzini, P., and M. Mariotti (2007): “Sequentially Rationalizable

Choice,” The American Economic Review, 5, 1824–1839.

Masatlioglu, Y., D. Nakajima, and E. Y. Ozbay (2012): “Revealed

Attention,” The American Economic Review, 102, 2183–2205.

Masatlioglu, Y., and E. A. Ok (2005): “Rational Choice with Status

Quo Bias,” Journal of Economic Theory, 121, 1–29.

Masatlioglu, Y., and E. A. Ok (2014): “A Canonical Model of Choice

with Initial Endowments,” The Review of Economic Studies, 81(2), 851–883.

Morris, S. (2000): “Contagion,” The Review of Economic Studies, 67(1),

57–78.

Ok, E. A., P. Ortoleva, and G. Riella (2015): “Revealed (P) Reference

Theory,” The American Economic Review, 105(1), 299–321.

Pesendorfer, W. (1995): “Design innovation and fashion cycles,” The

American Economic Review, pp. 771–792.

Pettibone, J. C., and D. H. Wedell (2000): “Examining Models of

Nondominated Decoy Effects Across Judgment and Choice,” Organizational

Behavior and Human Decision Processes, 81, 300–328.

Pollak, R. A. (1970): “Habit Formation and Dynamic Demand Functions,”

The Journal of Political Economy, pp. 745–763.

(1976): “Interdependent Preferences,” The American Economic Re-

view, pp. 309–320.

28

Rader, T. (1963): “On the Existence of Utility Functions to Represent Pref-

erences,” Review of Economic Studies, 30, 229–232.

Ray, D. (2006): “Aspirations, Poverty, and Economic Change,” Understand-

ing Poverty, pp. 409–421.

Ray, D., and A. Robson (2012): “Status, Intertemporal Choice, and Risk-

Taking,” Econometrica, 80(4), 1505–1531.

Rubinstein, A., and Y. Salant (2006): “Two Comments on the Principle

of Revealed Preference,” Mimeo.

(2008): “(A,f): Choice with Frames,” Review of Economic Studies,

75, 1287–1296.

Rubinstein, A., and L. Zhou (1999): “Choice Problems with a ’Reference’

Point,” Mathematical Social Sciences, 37, 205–209.

Siegel, S. (1957): “Level of Aspiration and Decision Making,” Psychological

Review, 64, 253–262.

Spiegler, R. (2006): “Competition Over Agents with Boundedly Rational

Expectations,” Theoretical Economics, 1(2), 207–231.

(2011): Bounded Rationality and Industrial Organization. Oxford

University Press.

Stutzer, A. (2004): “The Role of Income Aspirations in Individual Happi-

ness,” Journal of Economic Behavior and Organization, 54, 89–109.

Van Heerde, H. J., P. S. Leeflang, and D. R. Wittink (2000): “The

Estimation of Pre- and Postpromotion Dips with Store-Level Scanner Data,”

Journal of Marketing Research, 37, 383–395.

Veblen, T. (1899): The Theory of the Leisure Class. Oxford University Press.

29

6 Appendix A

For all proofs below, a preference relation is a reflexive and transitive binary

relation.

Proof of Theorem 3.1

[⇒] Suppose C is a choice correspondence that satisfies Axioms 3.1-3.5. Define

a(Z) := C(Z,Z). By Axioms 3.1 and 3.4, there exists an anti-symmetric total

order � such that a(Z) = max(Z,�). Notice that x � y if and only if

{x} = C({x, y}, {x, y}). Moreover, by Axiom 3.5, � and a are continuous.

Suppose x, y, z ∈ X. Taking z as an aspiration, we define the aspiration

based preference as x �z y if x ∈ C({x, y}, {x, y, z}) and C({x, y, z}, {x, y, z}) =

{z}. Additionally, �∗ is defined on X2 as: (x, y) �∗ (z, y) if x �y z. Note

that both of these preferences are generally incomplete. To obtain a repre-

senting utility function, we could appeal to Levin (1983)’s Theorem 1, but for

readability, we instead rely upon Corollary 1 of Evren and Ok (2011). That

corollary requires �∗ to be closed-continuous and T := {(x, y) : y � x} to be

a locally compact separable metric space (which is an implication of T being a

compact metric space). Below we show that �∗ and T meet these conditions.

First, �∗ is closed-continuous. To see this, take any x, y, z, xn, yn, zn ∈X and n = 1, 2, . . . such that xn → x, yn → y, zn → z and (xn, yn) �∗

(zn, yn) for all n. By the definition of �∗, we have xn �yn zn, i.e. xn ∈C({xn, zn}, {xn, yn, zn}) and {yn} = C({xn, yn, zn}, {xn, yn, zn}). By Axioms

3.4 and 3.5, we have that x ∈ C({x, z}, {x, y, z}) and {y} = C({x, y, z}, {x, y, z}).Therefore x �y z ⇒ (x, y) �∗ (z, y).

Second, T is compact. To see this, notice that the Axioms 3.4 and 3.5 imply

that T is a closed subset of the compact metric space X2. Therefore, there

exists a continuous u : T → R so that if (x, y) �∗ (z, y), then u(x, y) ≥ u(z, y)

and likewise for the strict relation. Define the similarity measure d as

d(x, y) :=

u(y, y)− u(x, y) if y � x,

u(x, x)− u(y, x) otherwise

30

This similarity measure represents an agent’s choices, that is, C(S, Y ) =

argmins∈S d(s, a(Y )). For ease of notation, denote a(Y ) by a.

“C(S, Y ) ⊆ argmins∈S d(s, a(Y ))”

Take y ∈ C(S, Y ) and suppose y /∈ argmins∈S d(s, a(Y )). Then, there

exists z ∈ S s.t. d(z, a) < d(y, a). As a is the aspiration alternative in Y ,

a � z, y. By the definition of d, it is the case that u(z, a) > u(y, a). So, z �a ywhich implies that {z} = C({z, y}, {z, y, a}) and {a} = C({z, y, a}, {z, y, a}).Axiom 3.2 guarantees that y ∈ C({z, y}, Y ). By Axiom 3.3, it is the case that

C({z, y}, Y ) = C({z, y}, {z, y, a}). This is a contradiction as y ∈ C({z, y}, Y ) =

C({z, y}, {z, y, a}) = {z}.

“argmins∈S d(s, a(Y )) ⊆ C(S, Y )”

Consider z ∈ argmins∈S d(s, a(Y )) and assume z 6∈ C(S, Y ). As C is non-

empty valued, there must exist y ∈ C(S, Y ). By Axioms 3.2 and 3.3, it is

the case that {y} = C({z, y}, {a, z, y}). Then y �a z by the definition of �a.Thus, d(z, a) > d(y, a) by the definition of d, a contradiction.

By construction, d is continuous, reflexive and symmetric, but need not

satisfy the triangle inequality. Lemma 6.1 shows that there exists a continuous

metric d : X2 → R+ with the same distance ordering, that is,

d(x, y) ≤ d(z, w) if and only if d(x, y) ≤ d(z, w).

Thus, C(S, Y ) = argmins∈S

d(s, a(Y )) for all (S, Y ) ∈ C(X).

[⇐] Axioms 3.1, 3.2, 3.3, 3.4 all follow trivially.

To show Axiom 3.5, suppose xn ∈ C(Sn, Yn) for all n, where SnH→ S,

YnH→ Y and xn → x. Since Sn

H→ S, we know that ∀s ∈ S, there exist

sn ∈ Sn for n = 1, 2, . . . such that sn → s. By xn ∈ C(Sn, Yn), we have

that d(xn, a(Yn)) ≤ d(sn, a(Yn)) for all n. Continuity of a(·) guarantees that

a(Yn) → a(Y ) and taking limits of both sides gives d(x, y) ≤ d(s, y). Since

s was chosen arbitrarily, it is the case that x ∈ argmins∈S d(s, y) and thus

x ∈ C(S, Y ). 2

31

The previous theorem constructs a semimetric (a distance function without

the triangle inequality) and the following lemma shows that any semimetric

can be transformed into a metric while preserving its distance ordering over

pairs.

Definition: A function φ : X2 → R is a semimetric if

1. ∀x, y ∈ X, φ(x, y) = 0⇔ x = y (Reflexivity)

2. ∀x, y ∈ X, φ(x, y) = φ(y, x) (Symmetry)

3. ∀x, y ∈ X, φ(x, y) ≥ 0 (Non-Negativity)

Lemma 6.1 Take (X,D) to be a compact metric space and suppose φ : X2 →R+ is a semimetric. Then, there exists a continuous metric Dφ : X2 → R+

such that

φ(x, y) ≤ φ(z, w) if and only if Dφ(x, y) ≤ Dφ(z, w) for any x, y, z, w ∈ X.

An earlier version of this paper provided a complete proof of this lemma.

Around the same time, a mathematics paper, Ben Yaacov, Berenstein, and

Ferri (2011) independently proved an equivalent result (Theorem 2.8). Our

proof uses a different approach and elementary techniques, but to keep the

analysis to a minimum, we omit it from the current version of the paper. As

our proof is more readily approachable and applies directly to the current

framework, we provide it in an online appendix.

32

Good x

Good y

c

2 y*

1 yn

xn

x*

pnp

Figure 5: Continuity argument for Proposition 4.1.

Proof of Proposition 4.1:

Let c denote the consumed bundle with price p and income I. Denote

the Regions R1 = {x : x2 ≤ I, U(x) > U(c), (x − c) · (p,−1) < 0} and

R2 = {x : x2 ≤ I, U(x) > U(c), (x − c) · (p,−1) > 0}. Take a sequence

of prices less than p so that p1, p2 . . . → 0. Let xn = C(B(I, pN)). For

each n, choose yn in R1 so that pn · yn = I and so that limnyn = y∗ is in

R1 with y∗2 = I. Suppose that xn ∈ R2 infinitely often. Then, passing to

that subsequence, by compactness limn→∞

xn = x∗ exists and xn ≺ yn for all n.

Therefore, continuity implies that x∗ - y∗, but monotonicity stipulates that

y∗ is strictly preferred to all members of R2, including x∗, a contradiction.

Therefore, xn must eventually lie in R1 and then any such pn represents a

sales price which will have consumption in R1 and therefore a positive impact

on x consumption after the sale ends. 2

Proof of Proposition 4.2:

Reorder the agents according to the aspirational preference so that agent

1 has the most preferred alternative of all agents according to the aspirational

preferences, agent 2 has the most preferred alternative of all agents but 1

and agent i has the most preferred alternative of agents i, . . . , n. Call these

aspirational alternatives a1, . . . , an.

33

Regardless of the network structure, agent 1 always chooses a1. Agent

2 chooses a2 if she is not connected to agent 1 and otherwise chooses x2 =

C(S2, S2∪a1). In general, agent j chooses xj = C(Sj, Yj) where Yj = Sj∪{xi :

i ≤ j, (i, j) ∈ Γ}. Additionally, notice by the ordering of the agents that

C(Sj, Yj) = C(Sj, Zj) where Zi = Si ∪ {xj : (i, j) ∈ Γ} and there is no

circularity in the choices C(Sj, Yj). Thus, having the agents choose in order

provides an algorithm for finding a steady state, thus establishing existence.

As for uniqueness, consider another steady state (x′1, . . . , x′n) that differs

from the one defined above. Let i denote the minimal index such that xi 6= x′i.

Now, it must be the case that xi = C(Si, Zi) and x′i = C(Si, Z′i). But, as

before, recall that Zi can be intersected with Si ∪ {x1, . . . , xi−1} and there-

fore the minimality of i provides that both of the above choices are equal to

C(Si, Zi ∩ (Si ∪ {x1, . . . , xi−1})) = C(Si, Z′i ∩ (Si ∪ {x1, . . . , xi−1})) because

Zi ∩ (Si ∪ {x1, . . . , xi−1}) = Z ′i ∩ (Si ∪ {x1, . . . , xi−1}). 2

Proof of Proposition 4.3:

Define a state ω to be a linear order and a network. Every state is equally

likely. WLOG, reorder the agents as before so that max(S1, >) > max(S2, >

) > . . . > max(SN , >).

In a steady state, we say that an agent j is potentially influenced by another

agent i if (i, j) ∈ Γ and i’s choice is ranked higher than j’s best available

option according to �. If an agent is potentially influenced by at least one

other agent, then we say that agent is influenced as his choice is determined

by that of another agent.5 For N ≥ 8, define ZN to be the event that agents

8, . . . , N are influenced.

When there are exactly N agents, let CN be the event that all agents choose

the same color good. It suffices to show that infN≥1

Pr(CN) > 0.

Lemma: infN≥8

Pr(ZN) > 0

It suffices to show that supN≥8

Pr(ZCN) < 1. We make the following definitions,

5He may be connected to multiple agents who choose superior alternatives, and thus foreach pair, we say that the agent is potentially influenced, but certainly one such agent willinfluence him.

34

where k > i:

Tk = {w : Agent k is not influenced}

Ti,k = {w : Agent k is not potentially influenced by i}

The following equation show that an upper bound on Pr(Ti) may provide

an upper bound on Pr(ZCN):

ZCN =

N⋃i=8

Ti ⇒ Pr(ZCN) = Pr

(N⋃i=8

Ti

)≤

N∑i=8

Pr(Ti)

Claim 1: Pr(Tk) <

(3

4

)k−1Pr(Tk) = Pr

(k−1⋂j=1

Tj,k

)=

k−1∏i=1

Pr

(Ti,k :

i−1⋂j=1

Tj,k

)

≤k−1∏i=1

(Pr(Agent k is not connected to agent i :

i−1⋂j=1

Tj,k)

+ Pr(Agent k is connected to agent i AND agent i

is not choosing her maximal alternative :i−1⋂j=1

Tj,k)

)

=k−1∏i=1

(1

2+

1

2Pr(Agent i is not choosing her maximal alternative :

i−1⋂j=1

Tj,k)

)

<k−1∏i=1

3

4=

(3

4

)k−1To see the last inequality, notice that for any event in which agent i does

not choose her maximal alternative, there is another event (found by switching

the order of her two alternatives in the preference ranking) where that agent

chooses her maximal alternative. As this creates an injection from one event

into its complement and therefore its probability must be less than one half.

Thus, Pr(ZCN) ≤

N∑i=8

Pr(Ti) =N∑i=8

(3

4

)i<

∞∑i=8

(3

4

)i< 0.6 < 1

2

35

Let D7 be the event that agents 1, . . . , 7 all choose the same color good

and that it is maximally ranked for all of them.

Notice that Pr(CN) ≥ Pr(D7∩ZN). Thus, the proposition will be finished

by showing Pr(D7 ∩ ZN) ≥ Pr(ZN)/26 > 0 as Pr(ZN) has a positive uniform

lower bound.

To see this, consider the following mapping from ZN to D7 ∩ ZN : WLOG

assume that r1 � b1. For n = 2, . . . , 7, if bn � rn then switch their rankings

in the ordering �. Notice that if it is the case that all agents from 8, . . . , N

are being influenced because they are connected to other agents who choose

superior options that this remains the case. Thus, the above map goes from

D7 ∩ ZN to ZN .

Next, notice that the inverse image of each element in the range has size at

most 26. Thus, it must be that the domain is of size at most 26 of the image

space, that is, Pr(ZN) ≤ Pr(D7 ∩ ZN) ∗ 26. 2

Proof of Proposition 4.4

Denote the firms’ chosen prices as pRH , pRL , p

BH , p

BL . In a pure-strategy equi-

librium, each firm must earn a non-negative profit because a firm can always

guarantee zero profit by setting piL ≥ cL, piH ≥ cH .

Suppose that in a steady state that a low-quality good delivers maximal

utility, that is qL − pL ≥ qH − piH . Then, a firm can set piH = cH + pL − cL.

As qH − cH > qL − cL, it is the case that qH − (cH + pL − cL) > qL − pL

and the good is desired. Therefore, there is a profitable deviation to the price

p∗H = piH + ε as the high-quality good at this price is both desired by agents

and more profitable than the low-quality good.

Thus, from here on out, we suppose that for at least one of the high-quality

goods is utility maximal. Suppose that pRH < pBH . If pRH ≤ cH , then the red firm

must be making a non-negative profit with their low-quality good, and the red

firm can profitably increase pRH . If pRH > cH then as all agents aspire to the red

good, the blue firm can deviate and earn a profit by setting pBH = pRH , pBL = cL.

Therefore, in equilibrium pRH = pBH . As aspirations are split, each firm is a

monopolist over half of the low-quality market and will therefore both set the

optimal low-quality price pL = (cL + pH)/2. Finally, as firms must have no

36

profitable incentive to undercut each other, they must be earning zero profit.

The zero profit condition is (1 − pH)(pH − cH) + (pH − pL)(pL − cL) = 0.

Substituting in for the optimal pL and solving yields

p∗H =

(1

3

)(2 + 2cH − cL − 2

√c2H − cHcL − cH + c2L − cL + 1

)p∗L =

(1

3

)(1 + cH + cL −

√c2H − cHcL − cH + c2L − cL + 1

)As a check, one may see that cL < p∗L < p∗H < cH which leads to the observation

that the high-quality market is being subsidized by the agents purchasing from

the low-quality market.

To show (ii), as firms make zero profit in both the rational and aspirational

equilibria, we need only pay attention to the agent’s welfare. In the rational

equilibrium, this is WR = (1− cH)(qH − cH) + (cH − cL)(qL − cL) and in the

aspirational equilibrium, this isWA = (1−p∗H)(qH −p∗H) + (p∗H −p∗L)(qL−p∗L).

Adding the firms’ profit (1 − p∗H)(p∗H − cH) + (p∗H − p∗L)(p∗L − cL) = 0 to WA

yields WA = (1 − p∗H)(qH − cH) + (p∗H − p∗L)(qL − cL). This makes the two

welfare functions more easily comparable:

WA −WR = (cH − p∗H)(qH − cH) + (p∗H − cH + cL − p∗L)(qL − cL) ≥ 0

⇔ qH − cHqL − cL

≥ 1 +p∗L − cLcH − p∗H

= 1 +1 + cH − 2cL −

√c2H − cHcL − cH + c2L − cL + 1

cH − 2 + cL + 2√c2H − cHcL − cH + c2L − cL + 1

=2cH − 1− cL +

√c2H − cHcL − cH + c2L − cL + 1

cH − 2 + cL + 2√c2H − cHcL − cH + c2L − cL + 1

2

37

Appendix B -FOR ONLINE PUBLICATION

7 Extensions

The main representation relies upon five axioms, three of which captured our

notion of aspirations: WARP for Phantom-Free Problems, WARP Given Po-

tential Sets and Independence of Non-Aspirational Alternatives (3.1 - 3.3).

The other two, Single Aspiration Point (3.4) and Continuity (3.5), were tech-

nical and necessary to derive the specific representation. In this section, we

consider relaxations of the technical axioms. A preference relation is a reflexive

and transitive binary relation.

7.1 Relaxing the Continuity Axiom

Axiom 3.5 required upper hemi-continuity of the choice correspondence when

both the feasible and potential sets vary independently. We relax this assump-

tion by requiring upper hemi-continuity when (i) only the feasible sets vary

and (ii) both the feasible and potential sets are the same and vary together.

Axiom 7.1 (Upper Hemi-Continuity Given Aspiration) For any (S, Y ) ∈C(X) and x, x1, x2, . . . ∈ Y with xn → x, if xn ∈ c(S ∪ {xn}, Y ) for each

n, then x ∈ C(S ∪ {x}, Y ).

Axiom 7.2 (Upper Hemi-Continuity of the Aspiration Preference) For any

Y, Yn ∈ X , xn ∈ X, n = 1, 2, . . . such that YnH→ Y and xn → x,

if xn ∈ C(Yn, Yn) for each n, then x ∈ C(Y, Y ).

The main representation given in Theorem 3.1 is based upon a stronger

continuity axiom than those above. The following theorem characterizes a sim-

ilar representation under the weaker continuity conditions and from a technical

point of view, it is much simpler.

38

Theorem 7.1 C is an aspirational choice correspondence that satisfies Ax-

ioms 3.4, 7.1 and 7.2 if and only if there exist

1. a continuous linear order �,

2. a metric d : X2 → R+, where d(·, a) is lower semi-continuous on L�(a)

such that

C(S, Y ) = argmins∈S

d(s, a(Y )) for all (S, Y ) ∈ C(X), (1)

where a(Y ) is the �-maximum element of Y .

The difference between the two representations is that in Theorem 3.1 the

metric is continuous whereas here it is lower semi-continuous on the appropri-

ate domain.

7.2 Multiple Aspiration Points

We now also relax the Single Aspiration Point Axiom. As C(Y, Y ) need no

longer be a singleton, we refer to it as the agent’s aspiration set. We will

characterize a choice procedure based upon a subjective notion of distance d

and an aggregator φ. When there is a single aspiration point, the procedure

coincides with our main theorem: the agent chooses the closest feasible alter-

natives to her aspiration according to d. However, when an agent aspires to

a set of alternatives, each feasible alternative is associated with a vector of

distances to each aspiration point. For each feasible alternative, the aggrega-

tor translates the vector of distances into a score and the agent chooses the

feasible alternative(s) with the minimal score.

This characterization relies on the following axiom:

Axiom 7.3 (Aggregation of Indifferences) For any Y ∈ X and x, z ∈ Y ,

if C({x, z}, {x, y, z}) = {x, z} ∀y ∈ C(Y, Y ), then C({x, z}, Y ) = {x, z}.

39

The above axiom stipulates that if an agent is indifferent between x and z

with respect to every alternative in her aspiration set, then she is indifferent

between x and z when considering that aspiration set as a whole. The axiom

provides a link between problems with a single aspiration point and choice

problems with multiple aspirations. Note that if the Single Aspiration Point

Axiom holds, then the above axiom is trivially satisfied.

A vector of distances ~d ∈ R|X| between an alternative x and a set Z ⊆ X

is defined as follows:

~dz(x, Z) :=

d(x, z) if z ∈ Z

∞ otherwise

The zth entry of the vector takes the value d(x, z) if z is in Z and ∞otherwise.

Definition 7.1 A function f : RX+ → R is single-agreeing if f(~v) = va for

all ~v = (vx)x∈X such that vb =∞ whenever b 6= a.

If there are several aspirations, then there are several distances to consider

and the agent uses an aggregator to assign a score. However, when there is

only one aspiration point, there is a single distance to consider and we require

that the aggregator coincide with the distance function in that case. The

single-agreeing property guarantees this.

Theorem 7.2 C is an aspirational choice correspondence that satisfies Ax-

ioms 7.1, 7.2 and 7.3 if and only if there exist

1. a continuous complete preference relation �,

2. a metric d : X2 → R+, where d(·, a) is lower semi continuous on L�(a)

and

3. a single-agreeing aggregator φ : RX+ → R+ where φ ◦ ~d(·,A(Y )) is lower

semi-continuous on L�(A(Y )) for any Y ∈ X ,

40

such that

C(S, Y ) = argmins∈S

φ(~d(s,A(Y ))) for all (S, Y ) ∈ C(X), (2)

where A(Y ) is the �-maximal elements of Y and φ ◦ ~d is such that

A(Y ) = argmins∈Y

φ(~d(s,A(Y ))) for any x ∈ X and Y ∈ X .

We interpret the above representation in the following manner. When a

decision maker is confronted with a choice problem (S, Y ), she first forms her

set of aspiration points A(Y ) by maximizing her aspiration preference � over

Y . For each feasible alternative s, she considers its distance to each aspiration,

giving rise to the vector ~d(s,A(Y )). She aggregates this vector and chooses

the alternative with the lowest score. The aggregator φ can be interpreted as

a measure of dissimilarity between alternatives and aspiration sets.

When the agent always has a single aspiration point, the single-agreeing

property implies that choices are made only according to the distance func-

tion as in the main representation. This is best illustrated with an example.

Consider an agent with a single aspiration a and three alternatives to choose

from: x, y, z. The generated distance vectors are then:d(x, a)

∞∞∞

,d(y, a)

∞∞∞

,d(z, a)

∞∞∞

and single-agreement requires that φ takes these vectors to d(x, a), d(y, a), and

d(z, a), respectively. Formally, φ(~d(s, a)) = d(s, a) and therefore minimizing

the former is equivalent to minimizing the latter. Put differently, the agent

uses her distance function d whenever she can and aggregates otherwise.

A wide range of aggregators are permitted, but desirable properties, such

as monotonicity, could be imposed through additional axioms:

41

Axiom 7.4 (Monotonicity) For any Y ∈ X and x, z ∈ Y ,

(i) x ∈ C({x, z}, {x, y, z}) for all y ∈ C(Y, Y ) implies x ∈ C({x, z}, Y ).

(ii) C({x, z}, {x, y, z}) = {x} for all y ∈ C(Y, Y ) implies C({x, z}, Y ) = {x}.

The representation makes clear that Axiom 7.3 is the minimal assumption

necessary. If an agent is indifferent between two alternatives with respect to

each aspiration in the aspiration set, then both of these alternatives generate

the same vector of distances with respect to that aspiration set and thus they

must be assigned the same aggregate score.

8 Additional Proofs

For the Appendix below, a preference relation is a reflexive and transitive

binary relation.

Proof of Lemma 6.1

If X is a singleton, then φ is the constant 0 function and D = φ satisfies

all of the above properties.

If X is not a singleton, then WLOG, assume maxx,y∈X

φ(x, y) = 1. This is

because maxx,y∈X

φ(x, y) exists since X2 is a compact space and φ is a continuous

function on that space. Therefore ψ :=φ

maxx,y∈X φ(x, y)is an order-preserving

transformation where maxx,y∈X

ψ(x, y) = 1.

The nature of the problem is that φ may fail the triangle inequality and

the goal is to find a continuous increasing transformation f on the distances

so that the triangle inequality will be satisfied. As in the figure, it could be

the case that φ(x, y), φ(x, z) are very small (but not equal to 0) and φ(y, z)

is close to 1 and it must be that f(φ(x, y)) + f(φ(x, z)) ≥ f(φ(y, z)).

To further analyze the problem, define the operator ∆ by ∆(ψ) : X3 → R3+

(x,y,z)→(ψ(x,y),ψ(x,z),ψ(y,z))

.

This map, for any triple outputs the length of the edges according to ψ of the

triangle defined by the triple. Additionally, notice that the operator ∆’s out-

put is ordered, so in general ∆ is not invariant to permutations of its inputs.

42

I[\

I[]

I\]

I\]I[\

I[]

Figure 6: Converting a triangle to satisfy the triangle inequality.

Let q denote a constant such that q ∈(

1

2, 1

).

Definition of D:

Now, we will define an increasing sequence of Dn such that their limit

defines D. The approach that we take fixes triangles with a “long” side first,

and fixes more and more triangles as n → ∞. Formally, at stage n; Dn will

be such that all triangles with longest side at least qn will satisfy the triangle

inequality according to Dn.

For the base case, let D0 = φ.

To proceed with the inductive construction of Dn, we define a few auxiliary

functions hi, gi, ki, and fi such that Di = fi ◦Di−1 .

• hi(a) := max{a, c− b : ∃x, y, z ∈ X3 s.t. (a, b, c) = ∆(Di−1)(x, y, z) and a ≤ b ≤ c

}.

Define hi(a) := a if 6 ∃(a, b, c) ∈ ∆(Di−1) where a ≤ b ≤ c.

• gi(a) := min

(hi(a),

qi−1

2

)on [0, qi). Define gi(q

i) = qi.

• ki : [0, qi] → [0, qi] to be the Upper Concave Envelope of gi. Recall

that the upper concave envelope of another function is the least concave

function that dominates the specified function.

• fi(a) :=

a a ∈ (qi, 1]

ki(a) a ∈ [0, qi]

43

The motivation for the preceding functions is as follows. hi guarantees

that hi(Di−1(x, y)) + hi(Di−1(y, z)) ≥ Di−1(x, z). So, hi is enough to fix an

unsatisfied triangle inequality if it is to be applied to only the two shorter sides

of the triangle. But hi presents two difficulties: i) it may be that one of the

two augmented sides may now be longer than the sum of the other two sides

and ii) hi cannot be applied to only two of the triangle’s sides. gi addresses the

first difficulty as gi(a) ≤ qi,∀a ≤ qi. Then, ki addresses the second difficulty

as it transforms gi into a continuous concave positive increasing function so

that ki(qi) = qi and ki(Di−1(x, y)) + ki(Di−1(y, z)) ≥ Di−1(x, z) whenever

Di−1(x, z) ≤ qi−1. Notice that this could not be done directly to hi because

hi(x) may be greater than qi. Finally, fi applies k to “small distances” and

leaves “large distances” unchanged.

Formally, we next prove that the following inductive properties hold true:

1. Di is continuous on X2.

2. ∀i ≥ 1 ∀x, y ∈ X, Di(x, y) ≥ Di−1(x, y).

3. ∀i, j ≥ 0, ∀x, y ∈ X, φ(x, y) ≥ qmin(i,j) ⇒ Di(x, y) = Dj(x, y).

4. If x, y, z ∈ X and Di−1(y, z) ≥ qi, then Di(y, z) ≤ Di(x, y) +Di(x, z).

5. ∀i ≥ 1, Di ∼X Di−1

6. ∀i ≥ 0, Di : X2 → R+ is symmetric, reflexive, and non-negative.

Proof of Properties 1,2,5:

Suppose (0, b, c) = ∆(Di−1)(x, y, z). Then 0 = Di−1(x, y) ⇒ x = y by the

reflexivity of Di−1.

Therefore hi(0) = maxDi−1(z, x)−Di−1(z, x) = 0.

So, gi(0) = min(0,qi−1

2) = 0 and gi(q

i) = qi.

Moreover, ∀z < qi, gi(z) ≤ qi−1

2= qi

(1

2q

)< qi.

44

Claim: hi is upper semi-continuous. Thus gi is upper semi-continuous as

well.

Consider an → a such that hi(an) → z. If it is the case that hi(an) = an

infinitely often, then z = a and by definition hi(a) ≥ a. On the other hand, if it

is the case that hi(an) > an infinitely often, then there exist triangles xn, yn, zn

such that Di−1(xn, yn) = an and Di−1(yn, zn)−Di−1(xn, zn) = hi(an). By the

compactness of X and the continuity of Di−1, one can pass to convergent

subsequences of xn, yn, zn and thus hi(an) = Di−1(yn, zn) − Di−1(xn, zn) →Di−1(y, z)−Di−1(x, z) ≤ hi(a) where the last inequality is becauseDi−1(x, z) =

a. The claim is proven.

Importantly, non-negativity of gi, hi, gi(0) = hi(0) = 0, and upper semi-

continuity of gi, hi guarantee full continuity of gi, hi at 0. So, it is the case that

ki is strictly increasing, concave, and continuous on [0, qi] where ki(0) = 0 and

ki(qi) = qi. Thus fi is strictly increasing, continuous on [0, 1] which implies

that Di ∼X Di−1 (Properties 1 & 5).

By the concavity of ki on [0, qi] and ki(x) = x for x ∈ {0, qi}, it is the

case that ki(x) ≥ x ∀x ∈ (0, qi). Therefore fi(a) ≥ a on [0, 1] which implies

∀x, y ∈ X, Di(x, y) ≥ Di−1(x, y) (Property 2).

Proof of Property 6:

Di(x, y) ≥ Di−1(x, y) ≥ 0 implies that Di is non-negative.

Di(x, y) = fi ◦ Di−1(x, y) = fi ◦ Di−1(y, x) = Di(y, x) where the middle

equality follows from the symmetry of Di−1.

Finally 0 = Di(x, y) ⇔ 0 = fi(Di−1(x, y)) ⇔ 0 = Di−1(x, y) ⇔ x = y

where the second implication comes from fi being a strictly increasing (hence

invertible) function on [0, 1] with fi(0) = 0 and the third implication is due to

the reflexivity of Di−1.

Proof of Property 3:

Let i < j. Then φ(x, y) = D0(x, y) ≤ . . . ≤ Di(x, y) ⇒ qi ≤ Di(x, y) =

Di+1(x, y) = . . . = Dj(x, y).

45

Proof of Property 4:

Take x, y, z such that Di−1(y, z) ≥ qi.

Case 1: Suppose Di−1(y, z) ≥ qi−1.

Then, by the inductive hypothesis, definition of Di and property 2, it is

the case that

Di(y, z) = Di−1(y, z) ≤ Di−1(x, y) +Di−1(x, z) ≤ Di(x, y) +Di(x, z)

Case 2: Suppose Di−1(x, y), Di−1(x, z), Di−1(y, z) ≥ qi and Di−1(y, z) <

qi−1.

Then Di−1(x, y) + Di−1(x, z) ≥ 2qi = 2q(qi−1) > qi−1 > Di−1(y, z) where

the second to last inequality holds because q > 1/2.

Case 3: Suppose qi ≤ Di−1(y, z) < qi−1 and WLOG both of the following

hold: Di−1(x, y) ≤ Di−1(x, z) and Di−1(x, y) < qi.

If hi(Di−1(x, y)) ≥ qi−1

2, then Di(x, y) = ki(Di−1(x, y)) ≥ gi(Di−1(x, y)) =

qi−1

2. By Property 5), it is the case that Di(x, z) ≥ Di(x, y) ≥ qi−1

2, hence

Di(x, y) +Di(x, z) ≥ 2

(qi−1

2

)= qi−1 > Di−1(y, z).

If hi(Di−1(x, y)) <qi−1

2, then hi(Di−1(x, y)) = gi(Di−1(x, y)) ≤ ki(Di−1(x, y)) =

Di(x, y). So Di(x, y) + Di(x, z) ≥ hi(Di−1(x, y)) + Di−1(x, z) ≥ Di−1(y, z) =

Di(y, z) where the last inequality is due to the definition of hi.

D is well-defined:

Define D(x, y) := limn→∞

Dn(x, y).

Notice that Dn ≤ Dn+1 ≤ 1 where the first inequality is due to property 2

and the last inequality is by definition.

Additionally, notice that D could instead be defined as D = f ◦ φ where

f = . . .◦f2◦f1. The function f is well-defined because fi(x) ≥ x, and fi(x) = x

if x ≥ qi. Thus, each x > 0 is touched by at most dlogq(x)e iterations of f and

fi(0) = 0⇒ f(0) = 0.

46

D is continuous:

By properties 3 and 5, it is the case that ||Di − Dj||∞ < qmin(i,j) which

implies that Di converges uniformly. Property 1 stipulates that each Di is

continuous and thus D is continuous because the uniform limit of a sequence

of continuous functions is continuous.

D(x,y) < D(z,w) iff φ(x,y) < φ(z,w):

If φ(x, y) < φ(z, w), then ∀i, it is the case that Di(x, y) < Di(z, w) by Prop-

erty 5. Moreover, if x 6= y, then ∃n such that qn < φ(x, y). By the definition of

D and property 3). one finds that D(x, y) = Dn(x, y) < Dn(z, w) = D(z, w).

On the other hand, if x = y, then ∀i, Di(x, y) = 0 by the definition of fi

and therefore D(x, y) = 0. Moreover, D(x, y) = 0 < φ(z, w) = D0(z, w) <

D1(z, w) < . . . < Di(z, w) < Di+1(z, w) < . . . < D(z, w). The above argument

holds exactly the same for the case where φ(x, y) ≤ φ(z, w).

D satisfies symmetry, reflexitvity and non-negativity:

This is immediate from properties 2, 6 and the definition of D.

D satisfies the triangle inequality:

Suppose y 6= z, then ∃n s.t. qn < φ(y, z). Then, by Property 3, qn <

Dn(y, z) and by the definition of D and Properties 3-5, it is the case that

D(y, z) = Dn(y, z) ≤ Dn(x, y) + Dn(x, z) ≤ D(x, y) + D(x, z). If y = z, then

D(y, z) = 0 ≤ D(x, y) +D(x, z) by the reflexivity and non-negativity of D. 2

2

Proof of Theorem 7.1

[⇒] Define a(Z) := C(Z,Z). By Axiom 3.1, we have that there exists a total

order � such that a(Z) = max(Z,�). By Axiom 3.4 we have � is a total anti-

symmetric order. Moreover, by Axiom 7.2, this order and a are continuous.

Consider S ⊆ L�(z). For any z ∈ X define Cz(S) := C(S, L�(z)). By Ax-

iom 3.2, we have that Cz is rationalizable on C(L�(z)) by a complete preference

relation �z. Alternatively, notice that x �z y iff x ∈ C({x, y}, {x, y, z}) and

C({x, y, z}, {x, y, z}) = z. This preference relation �z represents the agent’s

aspiration-dependent preference when he has aspiration z. We need to show

47

that �z satisfies the conditions for Rader’s Representation Theorem on L�(z).

X is compact, hence separable, so its topology has a countable base. Also, �zis upper hemi-continuous on L�(z) by Axiom 7.1.

Hence, Rader’s Representation Theorem guarantees that there exists

uz : L�(z) → R an upper semi-continuous function representing �z. WLOG,

take y � x. Define d(x, y) := −uy(x) + uy(y). Define other points sym-

metrically, i.e. if x � y, then d(x, y) := d(y, x). We point out here, that

since a(·) is a choice function, we have that x ∼ y ⇒ x = y and therefore

d(x, y) = d(x, x) = −ux(x) + ux(x) = 0. Finally let us notice that since uz is

upper semi-continuous on L�(z), we have that d(·, z) is lower semi-continuous

on L�(z).

While d satisfies reflexivity and symmetry, the triangle inequality may not

hold for d. However, if we define d = f ◦ d where f is the lower semi-continuous

transformation defined below, then d will satisfy the triangle inequality.

f(x) =

0 if x = 0,

1 +x

1 + xotherwise

Notice that d satisfies the triangle inequality for a rather trivial reason.

Consider x, y, z all distinct. Then d(x, y) + d(y, z) ≥ d(x, z) for the trivial

reason that 2 is a lower bound for the left hand side and an upper bound for

the right hand side. If x = z, then the right hand side is equal to 0 and we are

trivially done. Finally, if x = y or y = z, then both sides equal d(x, z) and we

have shown the triangle inequality in all cases. Lastly, notice that this is an

increasing transformation, so as long as d represents the preferences, so does

d.

Finally, for representability, we need that C(S, Y ) = argmins∈Sd(s, a(Y )) =

argmins∈S d(s, a(Y )). For notational ease, let a = a(Y ). Note that d(a, ·) is

lower semi-continuous on L�(a) and S is a compact subset of L�(a(Y )) and

hence the above argmin will exist.6

“C(S, Y ) ⊆ argmins∈S d(s, a(Y ))”

6By definition, a(Y ) = argmax(Y,�) and therefore L�(a(Y )) ⊇ Y ⊇ S.

48

Take y ∈ C(S, Y ), suppose ∃z ∈ S s.t. d(z, a) < d(y, a). Substitut-

ing, we get ua(z) > ua(y) ⇒ z �a y ⇒ {z} = C({z, y}, {z, y, a}) and

{a} = C({z, y, a}, {z, y, a}). Axiom 3.2 tells us that y ∈ C(S, Y ) ⇒ y ∈C({z, y}, Y ) and Axiom 3.3 yields y ∈ C({z, y}, Y ) = C({z, y}, {z, y, a}) =

{z}, a contradiction.

“argmins∈S d(s, a(Y )) ⊆ C(S, Y )”

Consider z ∈ argmins∈S d(s, a(Y )) and assume z 6∈ C(S, Y ), y ∈ C(S, Y ).

By Axioms 3.2 and 3.3 we must then have {y} = C({z, y}, {a, z, y})⇒ z ≺a y⇒ d(z, a) > d(y, a).

[⇐] Axioms 3.2, 3.1, 3.3, 3.4 all follow trivially. Axiom 7.2 follows from the

continuity of the aspiration map, a(·).To show Axiom 7.1, we first have that S, {xn} ⊆ L�(y). By continuity of

�, we therefore have that x ∈ L�(y). By our representation, we have that

∀s ∈ S, d(xn, y) ≤ d(s, y). Finally, by the lower semi-continuity of d(·, y)

on L�(y), we have that ∀s ∈ S, d(x, y) ≤ liminfn→∞d(xn, y) ≤ d(s, y) ⇒x ∈ argmins∈S∪xφ(s, y). 2

Proof of Theorem 7.2

[⇒] Step 1: We define the aspiration preference � and aspiration map A.

Define x � y if x ∈ C({x, y}, {x, y}). Since every pair {x, y} is compact,

we get that � is a total order. Moreover, by Axiom 7.2, the defined aspiration

choice correspondence A(Y ) := max(Y,�) is upper hemi-continuous and the

aspiration preference � is continuous.

As before, we will define preferences relative to a single aspiration point.

This definition will differ than the one given before because there may in fact

be multiple aspiration points for a given choice problem. However, in the case

of a single aspiration point, these definitions will be the same.

Step 2: We define the aspiration-based preferences �A and show that they

represent choice.

Definition: For any A ∈ X , we define

x �A y if x ∈ C({x, y}, {x, y} ∪ A) and A ⊆ C({x, y} ∪ A, {x, y} ∪ A).

49

Claim: For any (S, Y ) ∈ C(X), C(S, Y ) = max(S,�A) where A = A(Y ).

Proof : Let x ∈ C(S, Y ) and y ∈ Y . By Axioms 3.1, 3.2 and 3.3, we get

x ∈ C({x, y}, {x, y}∪A). By Axiom 3.3, we have A = C({x, y}∪A, {x, y}∪A).

Hence, by definition of �A, we obtain x �A y. As y is an arbitrary element

in Y , we get x ∈ max(S,�A). For the other inclusion, let x ∈ max(S,�A)

and suppose further that x /∈ C(S, Y ). Then there must exist y ∈ C(S, Y ).

By definition of �A, we have x ∈ C({x, y}, {x, y} ∪ A) and A ⊆ C({x, y} ∪A, {x, y}∪A). Applying Axioms 3.1 and 3.3 gives x ∈ C({x, y}, Y ). By Axiom

3.1, we get x ∈ C(S, Y ), which is a contradiction. 2

Step 3: We show that �z are upper semi-continuous and define the dis-

tance function.

Consider the case where xi �z y for all i and suppose xi → x. Then,

xi ∈ C({xi, y}, {xi, y, z}) and z ∈ C({xi, y, z}, {xi, y, z}). First, we notice that

z ∈ C({x, y, z}, {x, y, z}) by Axiom 7.2. Now, if xi ∈ C({xi, y, z}, {xi, y, z})happens infinitely often, then by Axiom 7.2, we have x ∈ C({x, y, z}, {x, y, z}),which implies that x ∈ C({x, y}, {x, y, z}) by Axiom 3.2. If xi /∈ C({xi, y, z}, {xi, y, z})infinitely often, then {z} = C({xi, y, z}, {xi, y, z}) infinitely often because if

y ∈ C({x, y, z}, {x, y, z}), then y ∼ z ⇒ y ∈ C({xi, y, z}, {xi, y, z}) and then

by Axiom 3.2 xi ∈ C({xi, y, z}, {xi, y, z}) for all i which is a contradiction

to xi /∈ C({xi, y, z}, {xi, y, z}) infinitely often. So, we can pass to a subse-

quence where it is only the case that {z} = C({xi, y, z}, {xi, y, z}). Finally,

either x ∈ C(∞∪i=1{xi} ∪ {y, z, x},

∞∪i=1{xi} ∪ {y, z, x}) or {z} = C(

∞∪i=1{xi} ∪

{y, z, x},∞∪i=1{xi}∪{y, z, x}). If the previous case holds, then we are done since

x ∈ C({x, y, z}, {x, y, z}) by Axiom 3.2 and 3.3. If the latter case holds, then

by Axiom 3.3, we get xi ∈ C({xi, y},∞∪i=1{xi} ∪ {y, z, x}) and Axiom 7.1 im-

plies that x ∈ C({x, y},∞∪i=1{xi} ∪ {y, z, x}). Then, by Axiom 3.3, we obtain

x ∈ C({x, y}, {x, y, z}).X is compact, hence separable and therefore L�(x) is separable (because

subspaces of separable spaces are separable). Upper semi-continuity of �xwas just shown above and �x is a complete preference relation. Therefore,

by Rader’s Theorem, there is a U : X2 → R+ such that U(·, z) is upper-semi

50

continuous on L�(z) and represents �z for any z ∈ X. Hence, U(x, z) ≥U(y, z) if and only if x �z y.7 Finally, define

d(x, y) =

0 if x = y,

1 x 6= y and x ∼ y

2− U(x, y)

1 + U(x, y)x ≺ y

2− U(y, x)

1 + U(y, x)x � y

The above d is symmetric, reflexive, satisfies the ∆-inequality and d(·, y)

is lower semi-continuous on L�(y).

Step 4: We show that �A are upper semi-continuous and define a set-

based aggregator function.

Notice that we are only concerned with A such that ∀a, b ∈ A, a ∼ b. For

any Y , consider the situation where A = A(Y ) and we have a sequence xi → x

and a z such that ∀i, xi �A z. Suppose there exists a ∈ A such that z � a.

Then, it must be the case that xi � z � a and xi ∈ A(Y ∪ xi ∪ z). Then by

Axiom 7.2, we have x ∈ A(Y ∪x∪z) and x �A z. Otherwise, consider the case

where xi � a � z infinitely often. Then, again by Axiom 7.2, we have that

x � a and x �A z. Finally, suppose that A(Y ∪ {xi, z}) = A infinitely often.

Then, by Axiom 7.2, we have that A(Y ∪ x) = A or A ∪ x. In the first case,

we can apply Axiom 7.1 and in the second, by virtue of x being an aspiration

alternative, we have x �A z.

Now, for any set A of this type, �A is upper semi-continuous and satisfies

the other conditions for Rader’s Representation Theorem on L�(A) by analo-

gous arguments to those in the previous paragraph where A = {z}. Also, by

our last claim, we have uA representing C(·, Y ) whenever A(Y ) = A.

7Rader’s theorem does not guarantee non-negativity of U . If U takes negative values, wecan always consider eU instead of U , which is non-negative and order-preserving. Hence,WLOG, we assume non-negativity of U function.

51

Hence, Rader’s Representation Theorem guarantees that there exists

uA : L�(A)→ R+ an upper semi-continuous function representing �A.8

Suppose that A = C(Y, Y ) for some Y ∈ X and x, y ∈ Y . Define φ(·) as

follows:

φ((x,A)) =

d(x, y) A = {y}

d(x, y) |A| = 2, x, y ∈ A, x 6= y

1 x ∈ A, |A| > 2

5− 1

1 + uA(a)− uA(x)x /∈ A, a ∈ A and |A| 6= 1

Notice that uA(a) is constant across all a ∈ A, so the fourth case above

is well-defined. It can be checked that the above function is lower semi-

continuous due to the upper semi-continuity of uA and the fact that 0 ≤d ≤ 2 < 4 ≤ 5− 1

1 + uA(a)− uA(x).

Now, we have only defined φ for certain tuples (x,A). This is because only

certain choice problems arise. More formally we make the following definition.

Step 5: We show that if two alternatives generate the same distance

vectors, then they have the same set-based aggregate score.

Definition: Let CP = {(x,A) : x ∈ Y for some Y ∈ X such that A(Y ) = A}

Claim: For any (x,A), (y,B) ∈ CP , where A = A(Y ), B = A(Z), if~d(x,A) = ~d(y,B), then φ(x,A) = φ(y,B).

Proof : First, if ~d(x,A) = ~0 = ~d(y,B) then φ(x,A) = φ(y,B) = 0. Next,

we show that A = B. Suppose not. Then, WLOG, ∃z ∈ B\A. Since,

1z∈A = 0, 1z∈B = 1, it must be the case that y = z. So, there can be at

most one alternative in B\A. If B contains only one alternative, then we

are in the previous case, so, let’s take another alternative b ∈ B ∩ A. Now,

y = z ∼ b⇒ d(y, b) = 1. Therefore, it must be the case that d(x, b) = 1. Thus

x ∼ b. But, then 0 = d(x, x) = d(y, x) when we consider the (now known to

8For the non-negativity of uA, please refer to footnote 7.

52

be aspirational) alternative x which implies that y = x. But, now we have a

contradiction because z ∈ B\A and z = x ∈ A.

If x = a for some a ∈ A, then 0 = d(x, a) = d(y, a) ⇒ x = y. Otherwise,

x, y ≺ a and ~d(x,A) = ~d(y, A), means d(x, a) = d(y, a) for any a ∈ A. This

means that A({x, y, a}) = a and {x, y} = C({x, y}, {x, y, a}). If |A| = 1, then

φ(x,A) = d(x, a) = d(y, a) = φ(x,A) for A = {a}. Otherwise, by Axiom 7.3,

{x, y} = C({x, y}, A ∪ {x, y}) ⇒ x ∼A y ⇒ uA(x) = uA(y) and the last case

of φ applies, so φ(x,A) = φ(y, A). 2

Step 6: We construct a distance based aggregator and show that it satisfies

the single-agreeing property.

So, we now know that φ : CP → R and ~d : CP → RX and the equivalence

relations defined by the inverse image of ~d is a refinement of φ. So, by a

standard argument, there exists a φ : RX → R that makes the diagram above

commute.

�

���� ���

�

�

�

�

��

��

���

�� �

Figure 7: Commuting Graph

Therefore φ(~d(x,A)) = φ(x,A). The case when a vector has one finite

entry is if |A| ≤ 1. Then φ and hence φ is defined by the first case to agree

with the distance function. Thus, φ has the “single-agreement” property.

Step 7: We show that the constructed objects represent the choice corre-

spondence.

53

For representability, we must show that C(S, Y ) = argmins∈S

φ(~d(s,A(Y ))).

First, let us note that φ ◦ ~d is lower semi-continuous, S is compact and

S ⊆ L�(A(Y )). Hence the above argmin will exist.9 For notational ease, let

A = A(Y ).

“C(S, Y ) ⊆ argmins∈S

φ(~d(s,A(Y )))”

Take y ∈ C(S, Y ), suppose ∃z ∈ S s.t. φ(~d(z, A)) < φ(~d(y, A)). We

consider the following cases:

1. A = {x}. Then the above reduces to d(z, x) < d(y, x). But, then

y /∈ C(S, Y )

2. A = {z, y}. Then the above becomes d(z, y) < d(y, z)

3. A = {w, y}, w 6= z. Then the above becomes 4 < 5− 1

1 + uA(y)− uA(z)<

d(w, y) < 2

4. z ∈ A, y /∈ A, then y /∈ C(S, Y )

5. z, y ∈ A, |A| > 2, then 1 < 1

6. z /∈ A, y ∈ A, |A| > 2. Then the above becomes φ(~d(z, A)) < 1

7. y, z /∈ A, |A| > 2. Then the above becomes

5− 1

1 + uA(a)− uA(z)< 5− 1

1 + uA(a)− uA(y)⇒ uA(y) < uA(z)

“argmins∈Sφ(~d(s,A(Y ))) ⊆ C(S, Y )”

Consider z ∈ argmins∈Sφ(~d(s,A(Y ))) and assume z 6∈ C(S, Y ), y ∈ C(S, Y ).

We consider the following cases:

1. Suppose z ≺ a ∈ A, |A| > 1. Then, it must be y ≺ a ∈ A and

5 − 1

1 + uA(a)− uA(z)≤ 5 − 1

1 + uA(a)− uA(y)⇒ uA(y) ≤ uA(z) and

therefore y ∈ C(S, Y ) = max(S,�A)⇒ z ∈ max(S,�A) = C(S, Y ) 9By definition, recall A(Y ) = argmax(Y,�) and therefore L�(Y ) ⊇ Y ⊇ S.

54

2. Suppose z ∈ A, then by Axioms 3.1 and 3.3, z ∈ C(S, Y )

3. Suppose z /∈ A, |A| = {a}. Then we have d(z, a) ≤ d(y, a)⇒ z �a y and

since y was chosen generically, we have z ∈ max(S,�a)⇒ z ∈ C(S, Y )

[⇐] Axioms 3.1, 3.2, 3.3, 7.3 all follow trivially. Axiom 7.2 follows from the

continuity of the aspiration preference �.

To show Axiom 7.1, let us consider the situation xn ∈ C(S ∪ {xn}, Y ) and

xn, x ∈ Y and xn → x, we have ∀s ∈ S, φ(~d(xn,A(Y ))) ≤ φ(~d(s,A(Y ))). By

lower semi-continuity of φ(~d(·,A(Y ))) we have that φ(~d(x,A(Y ))) ≤lim infn→∞

φ(~d(xn,A(Y ))) ≤ φ(~d(s,A(Y ))). Therefore x ∈ C(S ∪ {x}, Y ). 2

55