Arrow's Impossibility Theorem

I'm taking much of this material from Luce and Raiff..

Updated on October 1, 2001
Updated Aug 29, 2005

Social Welfare Functions

Suppose that we have a society of M agents.  Each agent has a set of actions that it is considering, A1,A2, ..., AM.  When each agent chooses an action ai from Ai, a consequence results c(a1,a2,...,aM).  Since each agent has its own goal, each agent has its own set of preferences defined over the set of possible consequences.  A social welfare function takes these preferences (and sometimes, utilities defined over these consequences) defined over the set of consequences and creates a preference pattern (and sometimes a utility) for the entire society.  This societal preference pattern or corresponding societal utility is called the social welfare, and a social welfare function is the function that maps individual preferences into the social welfare.  In discussing social welfare, we'll ignore actions and how consequences depend on actions.  Instead, we'll talk about a set of N  consequences, and denote these consequences {c1,c2, ... , cN}.

We should look at an example to get a more clear idea about what is going on.  Suppose that we have three agents, Bart, Lisa, and Maggie.  Suppose further that there are three consequences, and that the three agents have ordered the consequences according to their own preferences.
 

Preferences for the three agents.

Bart
Lisa
Maggie
Top
c1 c2 c1
Middle
c2 c1 c3
Bottom
c3 c3 c2

In the table, for example, Bart's favorite consequence is c1, followed by c2, followed by c3.  A social welfare function takes these preferences and creates an ordering for the entire society.  One possible ordering is that society prefers c1 to c2, and c2 to c3 (since two out of three agents have these preferences).  Then the preferences for each agent stay the same, but the preference for society is given by

Table 1: 
Preferences for the three agents, plus the preference due to social welfare considerations.

Bart
Lisa
Maggie
Society
Top
c1
c2
c1
c1
Middle
c2
c1
c3
c2
Bottom
c3
c3
c2
c3

We have constructed this social welfare function by balancing Bart's preferences with Lisa's, Lisa's with Maggie's, and Maggie's with Bart's.

If actions are involved and if society has the power to enforce its preferences than Bart, Lisa, and Maggie must choose actions the generate consequence c1.  In a way, society dictates what each agent should do.  This kind of feels funny to us because we like to think that we each can choose what we like.  We often don't like it when society (via the government, for example) tells us what to do, but it gets worse.  We are going to list a bunch of properties that any fair social welfare function should satisfy, perhaps one of the most desirable is that no individual should be able to dictate what is good for all of society.  What we'll find out is that we cannot simultaneoutly satisfy all of the properties because the properties cause our social welfare function to be dictated by a single individual.

It is important to be clear about what a social welfare function must do.  It must be able to look at the individual preference pattern for each individual, and combine the individual preference patterns into a preference pattern for society.  To get an idea of the complexity of defining such a social welfare function, consider Table 1.  This table represents a tiny portion of the social welfare function.  The table only tells us what the preference pattern for society is when Bart likes c1>c2>c3, Lisa likes c2>c1>c3, and Maggie likes c1>c3>c2.   The complete social welfare function will include another table for when Bart likes c1>c2>c3 and Lisa likes c2>c1>c3, but Maggie likes c1>c2>c3. 

Since, for three choices {c1,c2,c3}, there are 6 possible preference patterns for each individual, there are a toal of 6*6*6 possible tables of preference patterns like Table 1, and the social welfare function must unambiguously define the preference pattern for society for each of the 6*6*6 possible combinations of individual preference patterns.

We will use the term profile to mean a particular set of individual preference patterns.  So, Table 1 is one particular profile from the set of 6*6*6 possible profiles.


Desirable Properties of a Social Welfare Function

We can define a huge number of social welfare functions for any set of M-agent preference patterns.  Clearly, it is desirable to restrict the kind of functions that we will consider.  We now list six properties that our social welfare function (denoted by F) should satisfy.  After stating these properties, we'll give examples of the ones that need more clarification.
  1. This first condition restricts the nature of the social preference structure.
    1. The social preference ordering that results from applying the social welfare function should be defined for every pair of consequences.
    2. The social preference ordering that results from applying the social welfare function should be transitive; i.e., xPy & yPz ==> xPz.
  2. This condition deals with the size of societies, the domain of social welfare functions, and the number of consequences.
    1. The number of elements in the set of consequences is greater than or equal to three.  C = the set of choices, so |C|>=3;
    2. There are at least two agents in our society.  M denotes the set of all agents, so |M|>=2;
    3. The social welfare function F is defined for all possible profiles of individual orderings.
  3. (This condition is known as the positive association of social and individual values.) If an individual promotes (makes more preferred than it used to be) a consequence ci then the social preference has the following characteristic:
    1. For any pair of consequences ck and cm where ck is socially preferred to cm (k and m different from i) before ci is promoted by an individual, then ck is still socially preferred to cm after the promotion.
    2. Any social preference ordering that compares ci with any other alternative either remains unchanged or is changed in favor of ci; for example, if ci is socially preferred to ck before the an individual promotes ci, then ci cannot be less socially preferred than ck after the promotion.
  4. (This condition is known as the independence of irrelevant alternatives.)  Consider a set of consequences C1 that is a subset of a larger set of consequences C.  Suppose that I ignore the larger set of consequences for a minute, and ask each agent to create an ordering over the elements of C1.  I then apply my social welfare function and generate a preference pattern for society as a whole.  Suppose that I then include a consequence ck that is in C but not in C1 and ask each agent to put ck in its proper place in the preference ordering.  (This keeps the relative (pairwise) orderings of elements in C1 constant.) The ordering imposed by the social welfare function when it acts on the new set created from the union of C1 with {ck} should make sure that the pairwise orderings of C1 do not change.
  5. (This condition is known as citizen's sovereignty.)  For each pair of alternatives ci and cj, there is some profile of individual orderings such that society prefers ci to cj.
  6. (This condition is known as non-dictatorship.)  There is no individual agent with the property that whenever he or she prefers ci to cj (for any ci and cj) society does likewise, regardless of the preferences of other individuals.  This just means that a single individual cannot determine society's preferences no matter what others think.
We can now talk about these properties.  Since property 1 and 2 are trivial and since I've included a brief discussion about properties 5 and 6 in their respective definitions, we can concentrate on properties 3 and 4.

Let's begin by talking about property 3 --- the positive association of social and individual values.  Let's return to our example in Table 1, but we'll "promote" c2 such that it is more desirable.

Table 2:  Preferences for the three agents, plus the preference due to social welfare considerations after c2 has been promoted.

Bart
Lisa
Maggie
Society
Top
c1
c2
c1
c1
Middle
c2
c1
c2
c2
Bottom
c3
c3
c3
c3

The only difference between Table 1 and Table 2 is that Maggie now prefers c2 to c3 (Maggie promoted c2).  Since society preferred c2 to c3 before c2 was promoted, it better still prefer c2 to c3.

Now, let's turn to property 4 --- the independence of irrelevant alternatives.  The best way to illustrate this is to use an example of when it is violated.  Suppose that the consequences of the game are potential presidents: Bush, Gore, and Nader.  Suppose that rather than having a one-vote per person distribution, we instead give N points to the top choice of the individual, N-1 to the next choice, and so on, where N is the number of candidates.  We then add up the total points for each candidate, and the candidate with the highest number of points wins.
 

Preferences for the three agents and number of points vis-à-vis our social welfare function
 
Bart
Lisa
Maggie
Marge
Homer
Society
3
Gore
Bush
Gore
Bush
Bush
Bush: 11
2
Nader
Nader
Nader
Nader
Gore
Gore: 10
1
Bush
Gore
Bush
Gore
Nader
Nader: 9

Now suppose that we introduce a new candidate: Buchanan.  We use the same scoring scheme as before, but include Buchanan in the mix.
 
 

Preferences for the three agents and number of points vis-à-vis our social welfare function with one more candidate
  Bart
Lisa
Maggie
Marge
Homer
Society
4
Gore
Bush
Gore
Bush
Bush
Gore: 15
3
Buchanan
Nader
Nader
Nader
Gore
Bush: 14
2
Nader
Gore
Buchanan
Gore
Nader
Nader: 13
1
Bush
Buchanan
Bush
Buchanan
Buchanan
Buchanan: 9

Notice how Buchanan, the least popular candidate according to the table, tilted the vote in Gore's favor.  Property 4 states that we don't want this irrelevant alternative to affect the social choice.


The Impossibility Theorem

We begin by introducing a definition that we will use in the proof.  Consider the collection of M agents, and consider a subset S of these agents.  We say that the subset S of individuals is decisive for the ordered pair (c1,c2) if, whenever all the members of S prefer c1 to c2, society does likewise, regardless what the agents not in S have to say.  In other words, if S is decisive for (c1,c2)  then this coalition can always impose the preference c1 over c2 by having each member of the coalition express this preference.

It is important to be clear about what decisiveness means and what it does not mean.  Suppose that we have three agents: Bart, Lisa, and Maggie.  Suppose that these agents have three choices: {c1,c2,c3}.  Suppose that Bart and Lisa are decisive over (c1,c2) and are not decisive for any other pair of choices.  Now, consider the following sets of profiles for Bart, Lisa, and Maggie.
We state without proof the following theorem, which can be obtained by applying properties 2 through 5.

Pareto Optimality Theorem: The set of all individuals is decisive for every ordered pair (c1,c2).  In other words, if each individual prefers c1 to c2 then so does society.

It is useful to make one statement about how this proof would proceed.  Note that we cannot prove the Pareto Optimality Theorem using property 5 (citizen's sovereignty) alone.  All that citizen's sovereignty guarantees us is that, for a given pair of choices c1 and c2,  there is some set of individual profiles that will make society like c1 more than c2.  Returning to the three agent problem wtih Bart, Lisa, and Maggie, it may be that the precise set of profiles for which the social welfare function concludes c1>c2 is for Bart, Lisa, and Maggie to all express c3>c2>c1.  However, when we combine citizen's sovereignth with the positive association of social and individual values, we see that if everyone likes c1>c2 than so will society.

We now move to a useful equivalence.  This equivalence is the key to the proof so take some time to understand it.

Decisiveness Lemma: Assume condition 3.  A set S is decisive for (c1,c2) if and only if the following holds: society prefers c1 to c2 when all of the members of S prefer c1 to c2  and all other agents prefer c2 to c1.
Proof:  Since this is a standard if and only if proof, we'll take the standard approach.  We'll begin by showing that if a set S is decisive for (c1,c2) then society prefers c1 to c2 when all of the members of S prefer c1 to c2  and all other agents prefer c2 to c1.  Since, by the definition of decisiveness, the opinions of agents not in S have no influence over society's preference, it follows as a special case that when all members of S prefer c1 to c2  and all other agents prefer c2 to c1 then society prefers c1 to c2.

We'll now show that if society prefers c1 to c2 when all of the members of S prefer c1 to c2  and all other agents prefer c2 to c1, then set S is decisive for (c1,c2).  Every agent not in S prefers c2 to c1, so any changes in the individual preferences between these two consequences must promote c1.  But, by property 3, society must still prefer c1 to c2.  This means that no matter what agents not in S have to say, society still prefers c1 to c2 which means that the set S is decisive for (c1,c2).
Q.E.D.

Before we move onto the main theorem, we should probably talk about what this decisiveness lemma says.  Observe that decisiveness is a characteristic of the social welfare function itself and does not indicate anything about any individual agent's profile.  This means that if I am given the decisiveness property then the individuals in the decisive set can impose their will if they have the same preferences, but it does not mean that they will impose their will.  It is a statement of the potential for a group to impose their will if they choose to.

The first direction in the lemma says that if there is a decisive set then the agents in a decisive set can impose their preference pattern even if all other agents hold the exact opposite point of view; we get this just by the definition of decisiveness. 

The second direction is the more powerful direction.  It says that whenever we can show (a) society prefers c1 to c2, (b) a set S prefers c1 to c2, and (c) agents not in S prefer c2 to c1, then we can conclude that S is decisive for c1 for c2.  Let's be a bit more concrete about this.  Suppose that I have three agents, {Bart, Lisa, and Maggie}, and three choices, {c1,c2,c3}.  We have previously shown that the social welfare function for such a problem consists of 6*6*6 tables, one for each of the possible preference orderings of the agents.  Each of these tables shows what society would like if the agents had the preferences given in the table; it does not necessarily mean that the agents actually feel that way, just that if they did then we would know how society would order the choices.  Suppose that I start searching through the tables and found one table where (a) Bart and Lisa liked c1 more than c2, (b) Maggie liked c2 more than c1, and (c) society liked c1 more than c2.  By the decisiveness lemma, I would conclude that Bart and Lisa are decisive for c1 over c2, which means that I could conclude that every other table in which Bart and Lisa both had c1>c2 would imply that society also had c1>c2.  From that one table, I can conclude something about the social welfare function and, consequently, about a whole bunch of other tables in the definition of that function.

Thus, the second direction of the theorem means that if I can show that the social welfare function produces a social preference ordering that agrees with the preferences from a particular group even when everyone else disagrees, then I can say that that group is decisive.  Let me illustrate again using a slightly more abstract example.  In this example, there are three choices {A,B,C} and two groups (of possibly very many agents}, {S,T}.  Suppose that while I am searching through the definition of the social welfare function, I observe the following table: 


 S
Society
top
 A
C
A
middle
 B
B
C
bottom
 C
A
B

Note how S prefers A to B, T prefers B to A, and Society prefers A to B.  From this table alone and from the lemma we can conclude that S is decisive for A against B.  This means that I know that if I was considering a table where S liked A>B then I would be abel to conclude that society would also have A>B.



We can now turn attention to Arrow's impossibility theorem.

Arrow's Impossibility Theorem:  There does not exist a welfare function which possess the properties demanded by these conditions.  More specifically, if conditions 1 through 5 are satisfied then there exists an individual (a dictator) who is decisive for every ordered pair of alternatives.
Proof: We'll do this proof in eight steps.  There will be several mini-proofs and examples within the proof, so be careful to not lose the main thread of the proof.

set S=M; // We know that the set of all agents is decisive for everything.
int minsize=|S|; // The size of the smallest decisive set found so far.
choice c1,c2; // The choices that produce the smallest decisive set.
for (i=1; i<=|C|; i++)
{
    for (j=i; j<=|C|; j++)
    {
        // Search through each possible pairs of choices.
        // Find the smallest decisive set associated with that particular pair.
       foreach (s in 2M)  // searching from smallest set to biggest.  Skip the empty set.
       {
             foreach (profile in the Social Welfare Function set of tables)
             {
                   if (s is decisive according the decisiveness lemma in the profile for ci over cj);
                   {
                         if (|s| < minsize)
                         {
                               S=s;c1=ci;c2=cj;minsize=|S|;
                               break out of foreach (s in 2M) loop
                         }
                   }
             }
       }
    }
}
return (S, c1, c2);
// This pseudo code checks for all pairs of choices and for all subsets of agents
// whether the given subset of agents is decisive for that pair of choices according
// to the Decisiveness Lemma.  If so, it checks to see if the set is smaller than a previously
// identified decisive set.  If it is smaller, than it becomes the new minimally decisive set
// S, and the choices for which it is decisive are saved as c1 and c2.
//
// Note that I am performing an exhaustive set over choices and subgroups to find the
// smallest subgroup that is decisive for some pair of choices.  I do this by looking at
// profile tables and applying the decisiveness lemma.  The smallest possible decisive
// subgroup is returned along with the choices for which it is decisive.
Let's summarize what we've done to this point.  We have shown that a minimally decisive set exists and that it is nonempty (recall that by showing this, we did not assume anything about the preference patterns for the individuals).  We then asked ourselves what would happen if the individuals in society happened to have the profile shown in Table 3.  When these profiles hold, we find that the minimal decisive set is a singleton, consisting of individual j.  Since this is true for this profile and since the definition of a minimal decisive set accounts for all possible profiles, it follows that the minimal decisive set is a singleton.