You are on page 1of 14

Theoretical Economics 6 (2011), 95108 1555-7561/20110095

Transitive regret
Sushil Bikhchandani
Anderson School of Management, University of California, Los Angeles
Uzi Segal
Department of Economics, Boston College
Preferences may arise from regret, i.e., from comparisons with alternatives for-
gone by the decision maker. We ask whether regret-based behavior is consistent
with nonexpected utility theories of transitive choice and show that the answer
is no. If choices are governed by ex ante regret and rejoicing, then nonexpected
utility preferences must be intransitive.
Keywords. Regret, transitivity, nonexpected utility.
JEL classification. D81.
1. Introduction
Standard models of choice assume that decision makers act as if they maximize a prefer-
ence relation over sets of options and these preferences are assumed to be independent
of the environment. There are, however, good reasons to challenge this assumption.
Preferences may depend on the decision makers holding (reference point), on other
peoples holdings (envy), or on the choice set itself.
One such model is regret theory (Bell 1982 and Loomes and Sugden 1982). Accord-
ing to this theory the decision maker anticipates his future feelings about the choice he
is about to make and acts according to these feelings. This approach is natural when
the decision maker has to choose between two (or more) random variables. Once the
uncertainty is resolved, he will know what outcome he received, but also what outcome
he could have received had he chosen an alternative option. This comparison may lead
to rejoicingif his actual outcome is better than the alternativeor regret.
Formally, let A and Y be two random variables with money outcomes. Let (:, ,)
measure the regret or rejoicing a person feels when observing that he won : while the
alternative choice would have landed him,. Choosing A over Y thus leads, ex ante, to a
lottery +(A, Y) where the outcomes are (:, ,). Choice is based on regret and rejoicing
if there is a functional ! over regret/rejoice lotteries such that A is chosen over Y if and
only if ! (+(A, Y)) >0.
Sushil Bikhchandani: sbikhcha@anderson.ucla.edu
Uzi Segal: segalu@bc.edu
We thank Beth Allen, Kim Border, Eddie Dekel, Larry Epstein, Matt Jackson, Graham Loomes, Joel Sobel,
Bob Sugden, Peter Wakker, the editor, and the referees for helpful comments and suggestions.
Copyright 2011 Sushil Bikhchandani and Uzi Segal. Licensed under the Creative Commons Attribution-
NonCommercial License 3.0. Available at http://econtheory.org.
DOI: 10.3982/TE738
96 Bikhchandani and Segal Theoretical Economics 6 (2011)
The question we ask is simple: What functionals ! and regret/rejoice functions
are consistent with transitive choice? That is, when is it true that if ! (+(A, Y)) >0 and
! (+(Y, 7)) > 0, then ! (+(A, 7)) > 0 as well? If regret is separable across events, that
is, if ! (+(A, Y)) =

i
!
i
((:
i
, ,
i
), s
i
), then the possibility of having a violation of tran-
sitivity is well known (see Bell 1982, Loomes and Sugden 1982, and Fishburn 1989). In
fact, in that case, violations of transitivity must be observed unless (:, ,) =u(:) u(,),
which means that the original preferences are expected utility.
1
The main result of our
paper is that regret-based transitive choice implies expected utility and this conclusion
does not depend on ! being linear in probabilities or even separable across states.
To see why this result is not obvious, consider the following intuition. For equiprob-
able partition S
1
, . . . , S
n
, transitivity implies that for any vector of outcomes (:
1
, . . . , :
n
)
and any permutation of {1, . . . , n},
(:
1
, S
1
; . . . ; :
n
, S
n
) (:
(1)
, S
1
; . . . ; :
(n)
, S
n
)
(see Proposition 1 below). Separability of ! implies that the regret evaluation of
(:
i
, :
(i)
) in event S
i
does not depend on what happens in event S
]
, ] =i. Therefore, any
regret pair (:, ,) can be evaluated through a lottery and its permutation as above. With-
out separability this cannot be done, as the evaluation of the regret pair (:, ,) depends
on the rest of the lottery.
One can read the result of the paper in two different ways. It offers a necessary and
sufcient conditionfor a functional to be expectedutility without making any references
to mixture spaces (see Kreps 1988 for summary of terms and basic results). But the real
contribution is the impossibility result that shows that regret is inherently intransitive. If
so, then one must either conclude that (i) regret, despite its clear psychological appeal,
cannot be used in standard economic models; (ii) models of regret that are richer than
in Bell (1982) and Loomes and Sugden (1982) are necessaryfor example, as is done
in Sarver (2008) or by dening regret with respect to foregone distributions rather than
foregone outcomes (see Machina 1987 and Starmer 2000 for some steps in this direc-
tion); (iii) models of intransitive preferences must be incorporated into economics as in
Fishburn and LaValle (1988), Loomes and Sugden (1987), or Hayashi (2008).
2
The paper is organized as follows. The model and the main result are presented in
the next section. Section 3 offers an outline of the proof, while the details of the proof
appear in the Appendix.
2. The model and main result
Let L be the set of real nite-valued random variables over (S, l, P) with S = [0, 1], l
being the standard Borel algebra on S, P = being the Lebesgue measure, and the
set of outcomes being the bounded interval [

:, :]. The decision maker has a preference


relation over L. In the sequel, we denote events by S
i
and T
i
.
1
For this observation, see Sugden (2004, p. 739). We offer a formal proof of this claim in Lemma 7 below
as we are not aware of one in the literature.
2
See also Starmer (2000) for further references.
Theoretical Economics 6 (2011) Transitive regret 97
Definition 1. The continuous function : [

:, :] [

:, :] is a regret function if for


all :, (:, :) =0, (:, ,) is strictly increasing in : and strictly decreasing in ,.
If in some event A yields : and Y yields ,, then (:, ,) is a measure of the decision
makers ex post feelings (of regret if : - , or rejoicing if : > ,) about the choice of A
over Y. This leads to the next denition.
Definition 2. Let A, Y L, where A = (:
1
, S
1
; . . . ; :
n
, S
n
) and Y = (,
1
, S
1
; . . . ; ,
n
, S
n
).
The regret lottery evaluating the choice of A over Y is
+(A, Y) =((:
1
, ,
1
),
1
; . . . ; (:
n
, ,
n
),
n
),
where
i
= P(S
i
), i = 1, . . . , n. Denote the set of regret lotteries by R = {+(A, Y) :
A, Y L}.
For brevity we refer to and + as regret function and regret lottery, respectively,
even though they encompass both regret and rejoicing.
Definition 3. The preference relation is regret based if there is a regret function
and a continuous functional ! that is dened over regret lotteries such that for any
A, Y L,
A Y if and only if ! (+(A, Y)) 0.
The main result of this paper is the following.
Theorem 1. Let be a complete, transitive, continuous, and monotonic preference re-
lation over the set L of random variables. The relation is regret based if and only if it is
expected utility.
This theorem implies, in particular, the known result that the regret models of Bell
(1982), Loomes and Sugden (1982), and Sugden (1993) are intransitive.
3
We take this
result a step further and show that this intransitivity is not caused by separability across
events, but is the result of regret itself.
4
Recently, Sarver (2008) presented a nonexpected utility model of regret that is transi-
tive, but it departs fromthe standard regret model of Bell (1982) and Loomes and Sugden
(1982). In Sarvers model, the decision maker chooses between menus of lotteries and
a lottery from the selected menu. At the time these two choices are made, the decision
maker is uncertain about the utility of different outcomes. Later, after uncertainty is
3
An important exception is the case where the choice set consists of statistically independent random
variables, and for the two lotteries (:
1
,
1
; . . . ; :
n
,
n
) and (,
1
, q
1
; . . . ; ,
n
, q
n
), the probability of the regret
(:
i
, ,
]
) is
i
q
]
(see Machina 1987, pp. 138140 and Starmer 2000, pp. 355356). For example, Hong (1983)
weighted utility theory is consistent with this form of regret.
4
Guls (1991) model of disappointment is transitive and nonexpected utility. The comparison in this
model is between the outcome of a lottery and the lottery itself, rather than between possible outcomes of
a pair of lotteries.
98 Bikhchandani and Segal Theoretical Economics 6 (2011)
resolved, the decision maker may experience ex post regret if the selected lottery turns
out to be inferior to another lottery that is also in the menu he selected. This induces
a transitive, nonexpected utility preference relation over menus of lotteries in the initial
period. However, this is not inconsistent with Theorem 1. First, if menus are single-
tons, then Sarvers model reduces to expected utility. Second, the source of uncertainty
is different. In our model, the decision maker does not know which state of nature will
hold and, therefore, he does not know what outcome he will receive. In Sarvers model,
the decision maker does not know his future preferences and regret may emerge from
realizing that given his (now) known preferences, he chose the wrong option.
Theorem 1 is proved as follows. It is well known that expected utility is regret based
(with (:, ,) =u(:)u(,) and ! (+(A, Y)) =

i
(:
i
, ,
i
)). That any transitive regret-
based preferences must be expected utility is proved in a sequence of steps summarized
below.
Step 1. Preferences satisfy the equivalence condition (Loomes and Sugden 1982,
p. 818). That is, if A and Y have the same distribution, then A Y (Section 3.1,
Proposition 1).
Step 2. The indifference curve of ! through zero, {F: ! (F) =0}, is linear in probabili-
ties (Section 3.3, Lemmas 35).
Step 3. There exists ! as in Denition 3 that is linear in probabilities for all regret lot-
teries F (Section 3.4, Lemma 6).
Step 4. The preference relation is expected utility (Section 3.4, Lemma 7).
3. Proof of the theorem
3.1 Probabilistic equivalence
Whenpreferences are regret based, the decisionmaker cares about what events will hap-
pen as this will tell him what are the alternative outcomes he could have received had
he chosen differently. When the decision maker learns that the number 4 on a die yields
$100 under A and $150 under Y, the fact that these two outcomes are linked to the same
state of the world is important, but the state itself is not. Consequently, only the prob-
abilities of the underlying states are relevant for regret between A and Y. As long as
the probability of the number 1 is the same as that of 4, it makes no difference whether
the regret (100, 150) is obtained when the number is 1 or 4. This is why regret lotteries
are evaluated with respect to their probabilities and not with respect to the generating
events.
Proposition 1 shows that this observation, together with transitivity, has a signi-
cant implication to the evaluation of random variables. To see this, consider a box with
n balls, numbered 1, . . . , n. Draw one ball at random, and let A = (:
1
, S
1
; . . . ; :
n
, S
n
),
where S
i
is the event ball i is drawn. Let : {1, . . . , n} {1, . . . , n} be a permutation
of the n numbers and let (A) (:
(1)
, S
1
; . . . ; :
(n)
, S
n
). If A (A), then according
to the discussion in the last paragraph, it should also be the case that (A)
2
(A),

2
(A)
3
(A), . . . ,
n!1
(A)
n!
(A). By transitivity, we obtain that A
n!
(A) =A,
a contradiction.
Theoretical Economics 6 (2011) Transitive regret 99
For A L, let I
A
be the distribution of A, that is, I
A
(:) =P(A :).
Proposition 1 (Probabilistic equivalence). Let be a continuous and transitive regret-
based preference relation over L. For any two random variables A, Y L, if I
A
= I
Y
,
then A Y.
Loomes and Sugden (1987) and Fishburn and LaValle (1988) use cycles as above to
justify violations of transitivity. In Fishburn and LaValle (1988), a fair die is rolled and
payments are made according to the number shown. Consider the randomvariables A
1
and A
2
=(A
1
) given by
S
1
S
2
S
3
S
4
S
5
S
6
A
1
$1,000 $500 $600 $700 $800 $900
A
2
$900 $1,000 $500 $600 $700 $800
As in ve of six cases A
1
yields $100 more than A
2
, Fishburn and LaValle suggest that
preferring A
1
to A
2
is natural. But of course, using such a permutation ve more times
leads to a nontransitive cycle.
The converse of Proposition 1 is not true. As is demonstrated by the follow-
ing example, there are nontransitive regret-based preferences that satisfy probabilistic
equivalence.
Example 1. For two random variables A and Y, nd comonotonic A

and Y

with
the same distributions as A and Y. Formally, for A = (:
1
, S
1
; . . . ; :
n
, S
n
) and Y =
(,
1
, T
1
; . . . ; Y
n
, T
n
), nd A

= (:

1
, L
1
; . . . ; :

, L

) and Y

= (,

1
, L
1
; . . . ; ,

, L

) such that
:

1
:

, ,

1
,

, I
A
= I
A
, and I
Y
= I
Y
. Observe that A

and Y

depend
on both A and Y. Dene now A Y if and only if ! (A

, Y

) 0, where ! (A

, Y

) =

P(L
i
)(:

i
,

i
)
3
. In other words, is regret based with respect to the probability distri-
bution functions. As such, it satises probabilistic equivalence.
Let P(L
1
) =P(L
2
) =P(L
3
) =
1
3
. The random variables A, Y, 7 are given by
L
1
L
2
L
3
A 8 19 30
Y 9 20 28
7 10 18 29
Clearly ! (A, Y) =! (Y, 7) =! (7, A) =6, hence A Y, Y 7, but 7 A.
3.2 Preliminary results
We assume that outcomes are in a nite interval [

:, :]. Let

r = (

:, :) and r = ( :,

:).
By the continuity of the regret functional, -

r -0 - r -. As (:, ,) is continuous,
increasing in :, and decreasing in ,, it follows that the set of regret lotteries Rdened in
Denition 2 is the set of nite-valued lotteries with outcomes in the interval [

r, r]. The
following monotonicity properties of ! are inherited from the monotonicity of .
100 Bikhchandani and Segal Theoretical Economics 6 (2011)
Lemma 1. Let F and F

be two distinct regret lotteries such that F dominates F

by rst-
order stochastic dominance (FOSD).
(i) If ! (F) =0, then ! (F

) -0.
(ii) If ! (F

) =0, then ! (F) >0.


The next lemma permits a selection of regret lotteries that are skew symmetric in
regret and rejoicing.
Lemma 2. (i) If (:, ,) =(:

, ,

), then (,, :) =(,

, :

).
(ii) The equality (:, ,) =(,, :) is without loss of generality.
We will assume throughout that (:, ,) =(,, :) and that +(A, Y) =+(Y, A)
((,
1
, :
1
),
1
; . . . ; (,
n
, :
n
),
n
). Moreover,

r = r.
3.3 The indifference curve through zero is linear
A regret lottery F is generated by a permutation if there exists a random variable A =
(:
1
, S
1
; . . . ; :
n
, S
n
), P(S
i
) = 1,n, and a permutation of A such that +(A, (A)) = F.
By Proposition 1, if F is generated by a permutation, then ! (F) = 0. The next lemma
shows that the subset of {F: ! (F) =0} that is generated by permutations is convex.
Lemma 3. If F and F

are generated by permutations, then so is


1
2
F+
1
2
F

.
As F and F

are generated by permutations, we have ! (F) = ! (F

) = 0 and, by
Lemma 3, ! (
1
2
F +
1
2
F

) = 0. As is shown by the next example, one cannot guarantee


that every regret lottery F = (r
1
, 1,n; . . . ; r
n
, 1,n) such that ! (F) = 0 is generated by a
permutation.
Example 2. Consider an expected value maximizer whose choice set consists of ran-
dom variables with prizes in the interval [3, 3]. This individuals regret function is
(:, ,) =: , and he is indifferent between A and Y dened below, where P(S
i
) =0.2:
A = (3, S
1
; 3, S
2
; 1, S
3
; 1, S
4
; 1, S
5
)
Y = (3, S
1
; 3, S
2
; 3, S
3
; 3, S
4
; 3, S
5
).
As A Y, ! (+(A, Y)) =! (6, 0.2; 6, 0.2; 4, 0.2; 4, 0.2; 4, 0.2) =0. But there does not
exist a random variable

7 with outcomes in the interval [3, 3] and a permutation
such that +(A, Y) =+(

7, (

7)). To see why, observe that the rejoicing 6 must be gen-
erated by the outcomes 3 and 3. From outcome 3, only regret is possible, and as the
only regret level is 4, the outcome 3 must be paired with 1. From outcome 1, one
cannot generate rejoicing 6 or have regret 4.
5

5
If, instead, we had assumed that the set of outcomes was (, ), then any F = (r
1
, 1,n; . . . ; r
n
, 1,n)
such that ! (F) =0 would be generated by a permutation, leading to a simpler proof of Theorem 1.
Theoretical Economics 6 (2011) Transitive regret 101
The problem is that the outcomes in A and Y are far apart. However, as is shown by
the next example, one can nd in Example 2 a random variable 7 whose outcomes are
sufciently close to both A and Y such that A 7 Y, and the regret lotteries +(A, 7)
and +(7, Y) are generated by permutations.
Example 3. Using the notationof Example 2, let 7 =(0, S
1
; 0, S
2
; 1, S
3
; 1, S
4
; 1, S
5
). Thus
+(A, 7) =+(7, Y) =(3, 0.2; 3, 0.2; 2, 0.2; 2, 0.2; 2, 0.2).
Dene

7 = (3, S
1
; 0, S
2
; 3, S
3
; 1, S
4
; 1, S
5
)
(

7) = (0, S
1
; 3, S
2
; 1, S
3
; 1, S
4
; 3, S
5
).
Then +(

7, (

7)) =+(A, 7) =+(7, Y).
This idea is formalized below.
Lemma 4. Let A Y, where A = (:
1
, S
1
; . . . ; :
n
, S
n
), Y = (,
1
, S
1
; . . . ; ,
n
, S
n
), and
P(S
i
) = 1,n. Then there is a sequence A = 7
1
7
2
7
l
= Y such that for
every = 1, . . . , l 1, there is a regret lottery

7

and a permutation

so that
+(7

, 7
+1
) =+(

7

(

7

)).
Thus, even if a regret lottery F=(r
1
, 1,n; . . . ; r
n
, 1,n) with ! (F) =0 is not generated
by a permutation, one can nd a sequence of random variables 7
1
7
l
such that
each +(7

, 7
+1
) is generated by a permutation and F = +(7
1
, 7
l
). This is used to
prove that the set {F: ! (F) =0} is convex.
Lemma 5. If ! (F) =! (F

) =0, then ! (
1
2
F+
1
2
F

) =0.
3.4 V is linear in probabilities and is expected utility
The following lemma establishes that all indifference curves of ! are linear.
Lemma 6. (i) There is a function : : [ r, r] such that ! (F) 0 if and only if
E[:(F)] 0.
(ii) Moreover, : is strictly increasing with :(0) =0 and :((:, ,)) =:((,, :)) for all
:, ,.
We now use the function : to create a function u on outcomes that will turn out to
be the von NeumanMorgenstern utility claimed by Theorem 1.
Lemma 7. There exists an increasing function u: [

:, :] such that
:((:, ,)) =u(:) u(,).
102 Bikhchandani and Segal Theoretical Economics 6 (2011)
From the last two lemmas, we have for A = (:
1
, S
1
; . . . ; :
n
, S
n
) and Y = (,
1
, S
1
; . . . ;
,
n
, S
n
), where P(S
i
) =
i
,
A Y ! (+(A, Y)) 0

i
:((:
i
, ,
i
)) 0

i
[u(:
i
) u(,
i
)] 0
E[u(A)] E[u(Y)],
which is the claim of the theorem.
Appendix
Proof of Proposition 1. Let A = (:
1
, S
1
; . . . ; :
n
, S
n
) and Y = (,
1
, S

1
; . . . ; ,
n
, S

n
) be
such that I
A
=I
Y
.
Case 1. S
i
= S

i
and P(S
i
) = 1,n, i = 1, . . . , n. Then there is a permutation such that
Y = (A). Obviously, +(A, (A)) = +(
i
(A),
i+1
(A)). Hence, as there exists n n!
such that
n
(A) = A, it follows by transitivity that for all i, A
i
(A). In particular,
A Y.
Case 2. For all i, ], P(S
i
S

]
) is a rational number. Let N be a common denominator
of all these fractions. Random variables A and Y can now be written as in Case 1 with
equiprobable events T
1
, . . . , T
N
.
Case 3. There exist i and ], such that P(S
i
S

]
) is irrational. Any random variable
7 = (:
1
, T
1
; . . . ; :
n
, T
n
) is the limit of 7
l
= (:
l
1
, T
l
1
; . . . ; :
l
2
l
, T
l
2
l
), where for all l and ,
P(:
l

) =2
l
. This case follows by continuity from Case 2.

Proof of Lemma 1. Let F and F

be two regret lotteries. As usual, F dominates F

by
FOSD if and only if there is a list of probabilities
1
, . . . ,
n
adding up to 1 such that
F=(r
1
,
1
; . . . ; r
n
,
n
) and F

=(r

1
,
1
; . . . ; r

n
,
n
), and for all i, r
i
r

i
.
Fromthe continuity of , we knowthat for every r [

r, r] there exist :, , [

:, :] such
that r =(:, ,). Hence there are A, Y L such that +(A, Y) =F. By the continuity and
monotonicity of , we can nd A

and Y

such that :

i
:
i
, ,

i
,
i
, (:

i
, ,

i
) =r

i
for each
i, and +(A

, Y

) =F

. Either A strictly dominates A

by FOSD or Y

strictly dominates
Y by FOSD (or both). Monotonicity of implies that A A

and Y

Y with at least
one of these preferences being strict.
(i) If ! (F) = 0, then A Y. By transitivity, A

and hence ! (F

) =
! (+(A

, Y

)) -0.
Theoretical Economics 6 (2011) Transitive regret 103
(ii) If ! (F

) = 0, then A

. By transitivity, A Y and, therefore, ! (F) =


! (+(A, Y)) >0.

Proof of Lemma 2. (i) Let S
1
and S
2
be two disjoint events, where P(S
1
) = P(S
2
) =
0.5. Dene the lotteries A = (:, S
1
; ,, S
2
), Y = (,, S
1
; :, S
2
), A

= (:

, S
1
; ,

, S
2
), and
Y

=(,

, S
1
; :

, S
2
). Let r =(:, ,) =(:

, ,

). Then
+(A, Y) = (r, 0.5; (,, :), 0.5)
+(A

, Y

) = (r, 0.5; (,

, :

), 0.5).
By Proposition 1, A Y and A

; thus, we have ! (+(A, Y)) = ! (+(A

, Y

)) = 0.
But if (,, :) =(,

, :

), then +(A, Y) either dominates or is dominated by +(A

, Y

),
contradicting Lemma 1.
(ii) Recall that (:, :) =0. Let ] : [

r, r] [ r, r] be dened as
] (r) =

(,, :) if r -0 and : -, is such that (:, ,) =r


r if r 0.
By the rst part of this lemma, the value of ] (r) for r -0 does not depend on the choice
of :, , in the above denition; hence ] is well dened. Monotonicity of implies that ]
is strictly increasing. We can, therefore, dene
!

(r
1
,
1
; . . . ; r
n
,
n
) =! (]
1
(r
1
),
1
; . . . ; ]
1
(r
n
),
n
).
Let

(:, ,) =

(:, ,) if : ,
] ((:, ,)) if : -,.
Now
A Y ! (+(A, Y)) ! (+(Y, A))
!

(+

(A, Y)) !

(+

(Y, A)),
where +

(A, Y) is obtained from+(A, Y) by replacing (:, ,) with

(:, ,).
Proof of Lemma 3. In the sequel, random variables Q with n (not necessarily dis-
tinct) outcomes are of the form (q
1
, S
n
1
; . . . ; q
n
, S
n
n
) for some canonical partition where
P(S
n
i
) =1,n, i =1, . . . , n. For Q and Q

with noutcomes each, let


Q, Q

=(q
1
, S
2n
1
; . . . ; q
n
, S
2n
n
; q

1
, S
2n
n+1
; . . . ; q

n
, S
2n
2n
),
where P(S
2n
i
) =1,(2n).
Let F and F

be generated by permutations of A = (:
1
, S
1
; . . . ; :
n
, S
n
) and

of
Y = (,
1
, S

1
; . . . ; ,
n
, S

n
), respectively, where P(S
i
) = P(S

i
) = 1,n, i = 1, . . . , n. That is,
F =+(A, (A)) and F

=+(Y,

(Y)). (The assumption that A and Y are of the same


104 Bikhchandani and Segal Theoretical Economics 6 (2011)
length is without loss of generality.) Dene 7 =A, Y and

: {1, . . . , 2n} {1, . . . , 2n}


by

(i) =

(i) if i n

(i n) +n if i >n
to obtain +(7,

(7)) =+(A, Y,

A, Y) =
1
2
F+
1
2
F

.
Proof of Lemma 4. All random variables in this proof have n outcomes on the
equiprobable events S
1
, . . . , S
n
. For 7 = (:
1
, S
1
; . . . ; :
n
, S
n
) and 7

= (:

1
, S
1
; . . . ; :

n
, S
n
),
dene 7 7

=max
i
|:
i
:

i
|.
The proof follows from Claims 1 and 2.
Claim 1. Let A Y. For any > 0, there exist 7
1
, . . . , 7
l
such that A = 7
1

7
l
=Y and 7
1
7

, =2, . . . , l.
Proof. We construct the sequence 7
1
, . . . inductively. Suppose that A =Y and that we
have already dened A =7
1
7

such that 7
i1
7
i
, i =2, . . . , . If 7

=Y,
we are through. Otherwise, dene L

+
={i : :

i
>,
i
} and L

={i : :

i
-,
i
}. As 7

Y and
7

=Y, both L

+
and L

are nonempty. Let

+
= min
iL

+
{:

i
,
i
}

= min
iL

{,
i
:

i
}.
Dene ]

() such that 7

7
+1
() (:
+1
1
(), S
1
; . . . ; :
+1
n
(), S
n
), where
:
+1
i
() =

i
if i L

+
:

i
+]

() if i L
1

i
otherwise.
By continuity and monotonicity of , ]

() is well dened (for small ), continuous, and


increasing. Its inverse exists and is continuous. Dene

=min{,

+
, ]
1

)} and let
7
+1
=7
+1
(

). Note that 7
1
, . . . , 7
+1
satisfy the hypothesis of the claim.
If

= , then 7
+1
Y 7

Y . If

+
, then |L
+1
+
| |L

+
| 1. If

=]
1

), then |L
+1

| |L

| 1. Thus, this process terminates in a nite number of


steps with 7
l
=Y.
Claim 2. There exists
n
>0 such that if for all i, |r
i
| -
n
, then there exist a randomvari-
able

7 and a permutation such that F=(r
1
, 1,n; . . . ; r
n
, 1,n) satises F=+(

7, (

7)).
Proof. The domain of outcomes is [

:, :]. Let
:
1
=
: +

:
2

n
=
:

:
2n
=
:
1


:
n
=
: :
1
n
>0.
Theoretical Economics 6 (2011) Transitive regret 105
Thus, :
1
+n
n
= : and :
1
n
n
=

:.
The function is continuous on the compact segment [

:, :]; therefore, for any

n
>0, there exists
n
>0 such that |(:, ,)| -
n
implies |:,| -
n
. Thus, with |r
i
| -
n
we can construct

7 such that
S
1
S
2
S
3
S
4
S
n1
S
n

7 :
1
:
2
:
3
:
4
:
n1
:
n
(

7) :
2
:
3
:
4
:
5
:
n
:
1
+(

7, (

7)) r
1
r
2
r
3
r
4
r
n1
( :
n
, :
1
)
Outcome :
1
is chosen to be the midpoint between

: and :, and each :


+1
is cho-
sen so that ( :

, :
+1
) =r

, =1, 2, . . . , n 1. As |r

| -
n
, we have | :

:
+1
| -
n
and
each :

:, :]. As ! (F) = ! (+(



7, (

7))) = 0, it must be that ( :
n
, :
1
) = r
n
. Other-
wise, F either dominates or is dominated by +(

7, (

7)), contradicting Lemma 1. Thus,
F=+(

7, (

7)).
This completes the proof of Lemma 4.
Proof of Lemma 5. For F = (r
1
, 1,n; . . . ; r
n
, 1,n) and F

= (r

1
, 1,n; . . . ; r

n
, 1,n) such
that (F) = (F

) = 0, let A, Y, A

, Y

be such that +(A, Y) = F and +(A

, Y

) = F

.
By Lemma 4, there exist sequences A =7
1
7
l
=Y and A

=7

1
7

l
=Y

such that for all = 1, . . . , l 1 there exist



7

,

7

satisfying +(

7

(

7

)) =
+(7

, 7
+1
) and +(

7

(

7

)) = +(7

, 7

+1
).
6
Thus, for each = 1, 2, . . . , l 1, the
pair of regret lotteries +(7

, 7
+1
) and +(7

, 7

+1
) satises the hypothesis of Lemma 3.
Therefore,
!

1
2
+(7

, 7
+1
) +
1
2
+(7

, 7

+1
)

=0.
Note that
1
2
+(7

, 7
+1
) +
1
2
+(7

, 7

+1
) = +(7

, 7

, 7
+1
, 7

+1
), where , is de-
ned in the proof of Lemma 3. Consequently,
A, A

=7
1
, 7

1
7
l
, 7

l
=Y, Y

.
Hence
!

+(A, A

, Y, Y

=0,
but
+(A, A

, Y, Y

) =
1
2
F+
1
2
F

and we obtain ! (
1
2
F+
1
2
F

) =0.
As each A L is the limit of a sequence {A
l
}, where for each l, A
l
=(:
l
1
, 1,n
l
; . . . ;
:
l
n
l
, 1,n
l
), the lemma now follows by continuity for all F and F

such that ! (F) =


! (F

) =0.
6
We use the same l in both sequences without loss of generality, as the sequences may become station-
ary from a certain point on.
106 Bikhchandani and Segal Theoretical Economics 6 (2011)
Proof of Lemma 6. Recall that ! (
r
) > 0 > ! (
r
), where
i
is the constant lottery
yielding i.
(i) For a regret lottery F such that ! (F) > 0, let (F) be dened by ! ((F)F +
(1 (F))
r
) = 0, and for F such that ! (F) - 0, let (F) be dened by ! ((F)F +
(1 (F))
r
) = 0. By Lemma 1 and the continuity of ! , (F) is well dened and
(F) -1. Let

satisfy ! (

r
+(1

)
r
) =0.
We show rst that is a continuous function. Let F
l
F
0
and suppose that
(F
l
)

.
7
Suppose without loss of generality that for all l, ! (F
l
) 0. By the conti-
nuity of ! ,
! (

F
0
+(1

)
r
) =lim
l
!

(F
l
)F
l
+(1 (F
l
))
r

=0,
hence

=(F
0
).
Dene now
U(F) =

(F)

if ! (F) >0
0 if ! (F) =0
1

(F)
if ! (F) -0.
For F such that ! (F) = 0, (F) - 1; hence U(F) 0 if and only if ! (F) 0. The
continuity of () implies that U(F) is continuous. We show next that U is linear. That
is, for all F and F

, U(
1
2
F+
1
2
F

) =
1
2
U(F) +
1
2
U(F

).
By Lemma 5 and the continuity of ! we have the following conclusion.
Conclusion 1. If ! (F) =! (F

) =0, then for all [0, 1], ! (F+(1 )F

) =0.
For arbitrary regret lotteries F and F

, consider the three dimensional simplex A of


lotteries over F, F

,
r
,
r
. Take a linear transformation K of A such that K(
r
) =
(0, 0, 1), K(
r
) = (0, 0, (1

)), K(F) = (:

, ,

, :

), K(F

) = (:

, ,

, :

), and, by
Conclusion 1, ! (:, ,, :) =0 if and only if : =0. It follows that for : >0, (:, ,, :) solves
: (1 ) =0 (:, ,, :) =
1
: +1
and for : -0, (:, ,, :) solves
: +(1 )
1

=0 (:, ,, :) =
1

:
.
In both cases, U(:, ,, :) =

:.
Dene now a preference relation

on regret lotteries by F

if and only if
U(F) U(F

). Since U is continuous, so is

, and since U is linear,

satises the
independence axiom. Therefore, there is a function : such that U(F) 0 if and only if
E[:(F)] 0. The lemma follows since U(F) 0 if and only if ! (F) 0.
7
If (F
l
) does not have a limit, then we take a subsequence that has a limit.
Theoretical Economics 6 (2011) Transitive regret 107
(ii) Suppose that :() is not strictly increasing. Then there exists r
1
- r
2
such that
:(r
1
) :(r
2
). Take F=(r
1
,
1
; r
2
,
2
; . . . ; r
n
,
n
) such that ! (F) =0. The continuity of !
implies that such an F exists. Construct F

= (r
1
,
1
; r
2
,
2
+; . . . ; r
n
,
n
). Clearly
F

dominates F by FOSD, but 0 = ! (F) ! (F

), contradicting Lemma 1. The fact that


:(0) =0 follows from! (0, 1) =0.
Finally, let S
1
and S
2
be two disjoint events where P(S
1
) = P(S
2
) = 0.5. Dene
A = (:, S
1
; ,, S
2
) and Y (,, S
1
; :, S
2
). By Proposition 1, A Y. Thus :((:, ,)) =
:((,, :)).
Proof of Lemma 7. The following claim follows from a theorem in Aczl (1966) and is
mentioned, without an explicit proof, in Sugden (2004, p. 739).
Claim 3. If G(:, ,) + G(,, :) = G(:, :) for all : - , - :, then there exists a function
g: such that G(:, ,) =g(:) g(,).
Proof. Dene
H(:, ,) =

G(:, ,) if : -,
0 if : =,
G(,, :) if : >,.
It may be veried that for all :, ,, :,
H(:, ,) +H(,, :) =H(:, :).
Therefore, Aczl (1966, Theorem 1, p. 223) implies that there exists g: such that
H(:, ,) =g(:) g(,).
Select :
1
-:
2
-:
3
and , q >0, =q, +q -
1
3
. Dene lotteries A and Y as follows:
S
1
S
4
S
7
S
2
S
5
S
8
S
3
S
6
S
9
P(S
i
) q q q
1
3
q
1
3
q
1
3
q
A :
1
:
2
:
3
:
1
:
2
:
3
:
1
:
2
:
3
Y :
3
:
1
:
2
:
2
:
3
:
1
:
1
:
2
:
3
Proposition 1 implies A Y, as each of these lotteries gives :
1
, :
2
, and :
3
with
probability
1
3
each. Thus, ! (+(A, Y)) = 0 and, by Lemma 6, E[:(+(A, Y))] = 0. As
:((:, ,)) =:((,, :)) and :((:, :)) =0 (see Lemma 6), it follows that
[q ]:((:
1
, :
2
)) +[q ]:((:
2
, :
3
)) +[q]:((:
1
, :
3
)) =0.
Since = q, we obtain for all :
1
- :
2
- :
3
, :((:
1
, :
2
)) + :((:
2
, :
3
)) = :((:
1
, :
3
)).
By Claim 3, there exists a function u: such that :((:
1
, :
2
)) = u(:
1
) u(:
2
).
Monotonicity of u follows from the monotonicity of .
108 Bikhchandani and Segal Theoretical Economics 6 (2011)
References
Aczl, J. (1966), Lectures on Functional Equations and Their Applications. Academic
Press, New York. [107]
Bell, David E. (1982), Regret in decision making under uncertainty. Operations Re-
search, 30, 961981. [95, 96, 97]
Fishburn, Peter C. (1989), Non-transitive measurable utility for decision under uncer-
tainty. Journal of Mathematical Economics, 18, 187207. [96]
Fishburn, Peter C. and Irving H. LaValle (1988), Context-dependent choice with non-
linear and nontransitive preferences. Econometrica, 56, 12211239. [96, 99]
Gul, Faruk (1991), A theory of disappointment aversion. Econometrica, 59, 667686.
[97]
Hayashi, Takashi (2008), Regret aversion and opportunity dependence. Journal of Eco-
nomic Theory, 139, 242268. [96]
Hong, Chew Soo (1983), A generalization of the quasilinear mean with applications to
the measurement of income inequality and decision theory resolving the Allais para-
dox. Econometrica, 51, 10651092. [97]
Kreps, David M. (1988), Notes on the Theory of Choice. Westview Press, Boulder, Col-
orado. [96]
Loomes, Graham and Robert Sugden (1982), Regret theory: An alternative theory of
rational choice under uncertainty. Economic Journal, 92, 805824. [95, 96, 97, 98]
Loomes, Grahamand Robert Sugden (1987), Some implications of a more general form
of regret theory. Journal of Economic Theory, 41, 270287. [96, 99]
Machina, Mark J. (1987), Choice under uncertainty: Problems solved and unsolved.
Journal of Economic Perspectives, 1, 121154. [96, 97]
Sarver, Todd (2008), Anticipating regret: Why fewer options may be better. Economet-
rica, 76, 263305. [96, 97]
Starmer, Chris (2000), Developments in non-expected utility theory: The hunt for a
descriptive theory of choice under risk. Journal of Economic Literature, 38, 332382.
[96, 97]
Sugden, Robert (1993), An axiomatic foundation for regret theory. Journal of Economic
Theory, 60, 159180. [97]
Sugden, Robert (2004), Alternatives to expected utility: Foundations. In Handbook
of Utility Theory, Volume 2 (Salvador Barber, Peter J. Hammond, and Christian Seidl,
eds.), 685756, Kluwer, Dordrecht. [96, 107]
Submitted 2010-2-26. Final version accepted 2010-7-6. Available online 2010-7-6.