Professional Documents
Culture Documents
Singlepeaked 042
Singlepeaked 042
1. There is a group of agents (friends) who must share the job of watching
the barbecue. We will denote this set (group) of agents with N =
{1, 2, . . . , n}.
2. There is a fixed length of time T that the agents must take care of the
barbecue. This time interval is infinitely divisible.
3. At any time there will be exactly one agent taking care of the barbecue
(you need at least one person, and too many cooks will spoil the food).
4. Each agent only cares about how long they have to take care of the
barbecue, i.e., they do not care if they have to take care of the barbecue
from 11:00 to 12:00 or from 12:30 to 13:30, all that matters is the length
of the interval. Moreover for each agent there is an ideal duration and
given two two durations which are both less than or both greater than
this duration the one which is closer to the ideal duration is prefered
to the other. That is, each agent i has a preference relation Ri over the
interval [0, T ] such that: there is an ideal duration p(Ri ) (depending
on the preference relation of agent i) such that for any durations t, t′ ∈
[0, T ] if both t and t′ are less or if both are more than the ideal duration
p(Ri ), then agent i prefers the one which is closer to the ideal duration,
i.e., if t < t′ < p(Ri ) or p(Ri ) > t′ > t, then p(Ri ) Pi t′ and t′ Pi t.
We will call such preference relations single peaked (over the interval
[0, T ]).
1
In summary, the situation you are faced with can be modeled by a list
(N ; T ; (Ri )i∈N ), where N is the set of agents, T is a fixed time length that
must be shared between the agents, and for each i ∈ N , Ri is a single-peaked
preference relation of agent i over the set of possible durations that he can
be assigned, i.e., over [0, T ].
Given that you have modeled the situation, what you have to do next is
to solve the problem. What we mean by a solution in this case is to find a
list (t1 , t2 , . . . , tn ), where ti ∈ [0, T ] is the duration that agent i must take
care of the barbecue, such that t1 + t2 + · · · + tn = T .
Before we solve the barbecue problem lets look into a different problem.
Assume that, after the earthquake in the Marmara region, you are assigned
to the task of distributing the supplies that are sent to the region. You have
just received M tons of rice to be delivered to the region. What you have
to decide is how much to give each local station, which stores the supplies
in order to distribute to the local people. Assume that there are n local
stations. Each local station would like to receive a certain amount of rice,
lets call this their ideal amount. If they receive less than this there will be a
shortage of rice. If they receive more than this they will have some amount
of rice that they do not need. This will occupy valuable storage room that
could be used for other supplies. Also, the extra amount could have been
useful for other stations. Hence the local stations becomes worse off if you
give them more or less than their ideal amounts. Hence your problem is to
decide how to allocate a give amount M of rice to n local stations where
each possibly have different ideal amounts. Again you decide to model this
situation in order to solve your allocation problem. The following are the
essential ingredients of the problem:
2. There is a fixed amount M of the good and the good is infinitely divis-
ible.
3. For each agent there is an ideal amount and the closer the amount of
rice that they are given to this ideal amount, the better off the agent
is. That is, each agent i has a single-peaked preference relation over
the interval [0, M ].
In summary, the situation you are faced with can be modeled by a list
(N ; M ; (Ri )i∈N ), where N is the set of agents (local stations), M is a fixed
amount of rice that must be distributed between the agents, and for each
2
i ∈ N , Ri is a preference relation of agent i over the set of amounts of rice
that they can receive, i.e., over [0, M ]. We assume that each Ri , i ∈ N ,
is such that there is a p(Ri ) ∈ [0, M ], such that for any m, m′ ∈ [0, M ] if
m < m′ < p(Ri ) or p(Ri ) > m′ > m, then p(Ri ) Ri m′ and m′ Ri m.
But this is essentially the same problem that we were faced in our picnic
problem. Problems of this type will be called problems of allocating an
infinitely divisible good (time, rice, money, etc.) to a set of agents with
single-peaked preferences.
Some other similar situations are:
Note that in each of the situations given above, there is a set of agents,
N , and an amount, E, of some infinitely divisible good. This amount E is to
be divided among the agent in N . Hence each agent can receive an amount
between 0 and E. The agents has preference relations over the amount they
can receive, i.e., over the [0, E]. Each agent’s preference relation is such
3
that up to a certain point (which can be different for each agent) the agent
prefers receiving more to receiving less. After this point the agent prefers
receiving less to more. We will call such preference “single-peaked” preference
relations.
4
ui
xi
a c p(Ri ) d b e E
5
ui ui
xi xi
E E
ui
xi
E
6
ui
xi
E
ui
xi
E
ui
xi
E
ui
xi
E
7
ui
vi
ui
xi
a c p(Ri ) d b
ui
xi
xi x′i p(Ri ) x′′i E
8
If Ri is continuous, then U C(Ri , xi ) = [xi , ri (xi )] ∩ [0, E] if xi ≤ p(Ri ) and
U C(Ri , xi ) = [ri (xi ), xi ] ∩ [0, E] if xi ≥ p(Ri ).
The class of reflexive, complete, transitive, continuous, and single-peaked
preference relations will be denoted with Rsp .
∀i ∈ N yi Ri xi and ∃i ∈ N yi Pi xi .
If there exists such an y we will say that the allocation y Pareto dominates
the allocation x.
9
Definition 5 (Equal division rule) The equal division allocation for an
problem (N ; E; R) is the allocation in which each agent receives the same
amount, i.e., E/n. The rule which associates with each problem its equal
division allocation will be called the equal division rule, and will be denoted
with EA.
∀i ∈ N xi = max{0, p(Ri ) − λ}
P P
where λ ∈ R is such that xi = E (note that if p(Ri ) ≤ E, then λ ≤ 0)
will be called the equal distance allocation for (N ; E; R). The rule which
associates with each problem its equal distance allocation will be called equal
distance rule and denoted with ED.
10
1.3 Envy-free allocations
Definition 9 (Envy-free) Given a problem (N ; E; R), an allocation x is
envy-free if for any pair of agents i, j ∈ N we have xi Ri xj , i.e., if at x no
agent prefers some other agents share to his share. If for some pair i, j ∈ N
we have xj Pi xi , i.e., if agent i prefers agent j’s share to his share, we will
say that agent i envies agent j. An allocation rule is envy-free if it associates
envy-free allocations with each problem.
Proof : First we will show that thePproportional rule is Pareto efficient. Let
(N ; E; R) be a given problem. If p(Rj ) = 0 then, for any j ∈ N we have
p(Rj ) = 0 and P r(N ; E; R) = (E/n)j∈N . Thus, for any j ∈ N , p(Ri ) = 0 <
E/n = Fj (N ; E; R). Which, by Proposition 1, implies that F (N ; E; R) is a
Pareto efficient allocation. P P
Now we will
P consider the cases where p(R j ) > 0. If p(Rj ) ≤ E, then
we have E/ p(Rj ) ≥ 1. Therefore, for each i ∈ N ,
p(R )
P i E ≥ p(Ri ) .
p(Rj )
P P
If p(Rj ) ≥ E, then we have E/ p(Rj ) ≤ 1. Therefore, for each i ∈ N ,
p(R )
P i E ≤ p(Ri ) .
p(Rj )
Theorem 1 The uniform rule is the only peak-only allocation rule which is
Pareto efficient and envy-free.
11
Proof : Let F be a Pareto efficient and envy-free rule. Assume that F is not
equal to U . Then there Pexists a problem (N ; E; R)Psuch that F (N ; E; R) ̸=
U (N ; E; R). Assume p(Ri ) ≥ E (the case for ≤ E isP
p(Ri )P similar).
Let F (N ; E; R) = x and U (N ; E; R) = y. Since x ̸= y and xi = yi = E
there must be a pair k, l ∈ N such that xk < yk and xl > yl . We have two
cases to consider: xk < xl and xk ≥ xl
P
If xk < xl : y being Pareto efficient for (N ; E; R) and p(Ri ) ≥ E implies
that yk ≤ p(Rk ). This with xk < yk ≤ p(Rk ) implies that xk < p(Rk ).
If we also have xl < p(Rk ), then agent k would envy agent l. But this
contradicts with x being envy-free. Hence we must have xl > p(Rk ).
Let Rk′ be a single-peaked preference relation such that p(Rk′ ) = p(Rk )
and xl Rk′ xk (one can show that such a preference relations exists).
Consider the preference profile (R1 , . . . , Rk−1 , Rk′ , Rk+1 , . . . , Rn ) =
(Rk′ , R−k ). Note that the peaks of each preference relation in the profile
(Rk′ , R−k ) and the profile R are the same. This with F being peak-only
implies that F (N ; E; (Rk′ , R−k )) = F (N ; E; R) = x. But agent k envies
agent l at the allocation x when he is endowed with the preference Rk′
(note the construction of Rk′ ). This contradicts with F being envy-free.
Hence we can not have xk < xl .
i.e., F makes assignments in such a way that when the amount to be divided
changes all the agents are effected in the same way: the all become worse off
or they all become better off.
12
The equal division rule is not resource monotonic. Consider the case
where N = {1, 2}, E = 4, E ′ = 6, and R = (R1 , R2 ) is such that p(R1 ) = 1,
p(R2 ) = 3. Then EA(N ; E; R) = (2, 2) and EA(N ; E ′ ; R) = (3, 3). Note that
agent 1 prefers the allocation EA(N ; E; R) to the allocation EA(N ; E ′ ; R)
but agent 2 prefers the later to the former. Hence EA is not resource mono-
tonic.
Proof : Assume that there exists a resource monotonic and envy free allo-
cation rule F . Let N = {1, 2}. Let R = (R1 , R2 ) be as given in Figure 6.
E = 6 and E ′ = 20. Let x = F (N ; E; R) and x′ = F (N ; E ′ ; R).
E/2 E
3 6 12 14 E′
Since F is envy free the allocation x must be envy free. But then we
must have x1 ≤ x2 (otherwise agent 2 would envy agent 1 in the problem
(N ; E; R)). This with x1 + x2 = E = 6 implies x1 ≤ 3 and 3 ≤ x2 ≤ 6.
Now consider the allocation x′ which is envy free for the problem
(N ; E ′ ; R). Assume x′1 ≥ x′2 . This inequality with x′1 + x′2 = E ′ = 20
implies x′1 ≥ 20/2 = 10. But if this is the case agent 1 would envy agent 2
in the problem (N ; E ′ ; R), contradicting with the fact that x′ is envy free for
the given problem. Hence we must x′1 ≤ x′2 , which implies x′1 ≤ 10 ≤ x′2 .
If x′2 > 14, then x′1 = 20 − x′2 < 6 and agent 2 will envy agent 1. Thus we
must have x′2 ≤ 14, hence 10 ≤ x′2 ≤ 14 and 6 ≤ x′1 ≤ 10. These imply that
x1 P1 x′1 and x′2 R2 x2 , which contradicts with F being resource monotonic.
Hence we conclude that there is no resource monotonic allocation rule that
selects envy-free allocations for each problem. □
13
We will use Rfsp to denote the set of all preference profiles such that, for
each agent i there exists a positive amount xi which agent i is indifferent to
0, i.e., R ∈ Rfsp if and only if for each i ∈ N , there exists xi ∈ R++ such that
0 Ii xi .
14