Generalized Dictatorships

We now turn to the aggregation procedures calledgeneralized dictatorships. In our framework, an-ary aggregator is a generalized dictatorship if, on input anyn vec- tors from a domainD, always returns one of those vectors as its output. These ag- gregators are a natural generalization of the notion of dictatorial aggregators, in the sense that they select a possibly different "dictator" for each set ofnfeasible voting patters, instead of a single global one.

However, these aggregators were originally introduced by Cariani et al. [5] as rolling dictatorships, under the stronger requirement that the above property holds for anynvectors of{0, 1}^m. Under this requirement, a generalized dictatorship se- lects for each n-tuple of vectors from {0, 1}^m one of these vectors as its outcome (though not necessarily always the same one), hence the denomination "rolling" dic- tatorships. In that framework, Grandi and Endriss [13] showed that the class of generalized dictatorships coincides with the class of operators that are aggregators for every Boolean domainD⊆ {0, 1}^m.

The difference with our framework is that generalized dictatorships -and all ag- gregators in general- even though they are defined over{0, 1}^n×m, they are studied within the "rationality restrictions" of a given domain, in the sense that we are not interested in how they behave on irrational inputs, i.e. inputs outside of the domain.

In this section, we show that domains admitting aggregators which are not general- ized dictatorships are exactly the possibility domains (apart from some trivial cases), and are thus described by possibility integrity constraints, a result proved by Díaz et al. [9].

Definition 21. Let f = (f₁, . . . ,f_m)be an m-tuple of n-ary conservative functions. We say that f is ageneralized dictatorship for a domainD ⊆ {0, 1}^m, if, for any x¹, . . . ,xⁿ ∈ D, it holds that:

f(x¹, . . . ,xⁿ):= (f₁(x₁), . . . ,fm(xm))∈ {x¹, . . . ,xⁿ}. (3.1) Much like dictatorial functions, it is straightforward to observe that if f is a gen- eralized dictatorship forD, then it is also an aggregator forD.

The following example shows that the result of Grandi and Endriss [13, Theorem 16] does not hold in our setting, as it illustrates an aggregator that is a generalized dictatorship for one domain and not for another.

Let us first introduce some additional notation. In what follows, we will denote by ¯f ann-ary aggregator(f₁, . . . ,f_m)_{, where} f₁ = _{. . .} = f_m := f. Such aggregators are called systematic (see Section3.4).

Example 13. Consider the binary aggregator∧¯ = (∧,∧,∧)and the the Horn formulas:

φ₁₂ = (x₁∨ ¬x₂∨ ¬x₃)∧(¬x₁∨x₂∨ ¬x₃)∧(¬x₁∨ ¬x₂∨x₃)∧(¬x₁∨ ¬x₂∨ ¬x₃), and

3.2. Generalized Dictatorships 45

φ₁₃= (¬x₁∨x₂)∧(x₂∨ ¬x₃)∧(¬x₁∨ ¬x₂∨x₃). The set of satisfying assignments ofφ₁₂is:

Mod(φ₁₂) ={(0, 0, 0),(0, 0, 1),(0, 1, 0),(1, 0, 0)}.

By definition, Mod(φ₁₂)is a Horn domain and it thus admits the binary symmetric aggre- gator∧¯ = (∧,∧,∧). Furthermore,∧¯ is not a generalized dictatorship for Mod(φ₁₂), since

¯

∧((0, 0, 1),(0, 1, 0)) = (0, 0, 0)∈ {(/ 0, 0, 1),(0, 1, 0)}.

On the other hand,∧¯ is again an aggregator for the Horn domain:

Mod(φ₁₃) ={(0, 0, 0),(0, 1, 0),(0, 1, 1),(1, 1, 1)},

but, contrary to the previous case,∧¯ is a generalized dictatorship for Mod(φ₁₃), since it is easy to verify that for any x,y∈ Mod(φ₁₃),∧(¯ x,y)∈ {x,y}.

Finally, observe that(∧,∨,∨)is an aggregator for Mod(φ₁₃)that is not a generalized dictatorship. The latter claim follows from the fact that:

(∧,∨,∨)((0, 1, 0),(1, 1, 1)) = (0, 1, 1)∈ {(/ 0, 1, 0),(1, 1, 1)},

while the former is left to the reader. Thus, interestingly enough,φ₁₃describes a domain ad- mitting an aggregator that is not a generalized dictatorship, although it is not the aggregator that “corresponds” to the formula.

This means that deciding whether an aggregator is a generalized dictatorship depends on the domain in question. Therefore, generalized dictatorship as a feature is not necessarily an inbuilt property of an aggregator.

Despite the above, it is easy to see that a dictatorial aggregator (_prⁿ_i, . . . , prⁿ_i) is a generalized dictatorship for any D ⊆ {0, 1}^m, for all n ≥ 1 and for all i ∈ {1, . . . ,n}. Thus, trivially, every domain admits aggregators which are generalized dictatorships. The following result is also straightforward.

Lemma 7. Let D ⊆ {0, 1}^m, where|D| =2. Then, every aggregator f for D is a general- ized dictatorship.

Proof. Assume that D = {x,y}_{, where} x and y are distinct and that f is a n-ary aggregator for D that is not a generalized dictatorship for D. Then, it must hold either that f(x, . . . ,x) = yor f(y, . . . ,y) = x(the output of f must always bexory since f is an aggregator of{x,y}). Contradiction, since f is conservative.

Recall that a domain cannot have strictly less than two elements since we have assumed that it is not degenerate.

Again, the aim is to syntactically characterize domains that admit aggregators which are not generalized dictatorships. The following result shows that these do- mains are all the possibility domains with at least three elements, and are thus char- acterized by possibility integrity constraints.

Theorem 13(Dìaz et al. [9]). A domain D⊆ {0, 1}^m, with at least three elements, admits an aggregator that is not a generalized dictatorship if and only if it is a possibility domain.

Proof. The forward direction is obtained by the trivial fact that an aggregator that is not a generalized dictatorship is also non-dictatorial.

Now, suppose thatDis a possibility domain. Then it is either affine or it admits a binary non-dictatorial aggregator. We begin with the affine case. It is a known result that D ⊆ {0, 1}^m is affine if and only if it is closed under⊕, or, equivalently, if it admits the minority aggregator:

⊕¯ = (⊕, . . . ,⊕

| {z }

m-times

)

46 Chapter 3. Other forms of non-dictatorial aggregation- Characterizations of the corresponding domains

Claim 4. Let D⊆ {0, 1}^m be an affine domain. Then, theminorityaggregator:

⊕¯ = (⊕, . . . ,⊕

| {z }

m-times

)

is not a generalized dictatorship for D.

Proof. Let x,y,z ∈ D be three pairwise distinct vectors. Sincey 6= z, there exists a j ∈ {1, . . . ,m}such that y_j 6= z_j. It follows that y_j+z_j ≡ 1( _{mod 2}). This means that⊕(x_j,y_j,z_j) 6= x_j and thus that ¯⊕(x,y,z) 6= x. In the same way we show that

⊕(¯ x,y,z)∈ {/ x,y,z}, which is a contradiction, since ¯⊕is an aggregator forD.

Now, recall that if f = (f₁, . . . ,fm) is a binary non-dictatorial aggregator, then f_j ∈ {∧,∨, pr²₁, pr²₂}, j = 1, . . . ,m. If f_j ∈ {∧,∨} for all j ∈ {1, . . . ,m}, we call f symmetric, whereas if f_j ∈ {pr²₁, pr²₂}for all j ∈ {1, . . . ,m}, we call f a projection aggregator.

Claim 5. Suppose D⊆ {0, 1}^madmits a binary non-dictatorial non-symmetric aggregator f = (f₁, . . . ,f_m). Then f is not a generalized dictatorship.

Proof. Assume, to obtain a contradiction, that f is a generalized dictatorship forD and letx,y ∈ D. Then, f(x,y) := z ∈ {x,y}. Assume thatz = x. The case where z=yis analogous.

Let J ⊆ {1, . . . ,m}such that f_j is symmetric, for all j ∈ J and f_j is a projection otherwise. Note thatJ 6= {1, . . . ,m}. Let alsoI ⊆ {1, . . . ,m} \J, such that f_i =pr²₂, for alli∈ I and f_i =pr²₁otherwise. If I 6=∅, then, for alli∈ I, it holds that:

y_i = pr²₂(x_i,y_i) = f_i(x_i,y_i) =z_i = x_i.

Sincex,ywere arbitrary, it follows thatD_i ={x_i}, for alli∈ I. Contradiction, since Dis non-degenerate.

If I = _{∅, then} f_j = _pr²_{1, for all} j ∈/ J. Note that in that case, J 6= _{∅, lest} f is dictatorial. Now, consider f(y,x):=w∈ {x,y}since f is a generalized dictatorship.

By the definition of f,w_j = z_j = x_j, for all j∈ J, andw_i = y_i, for alli∈/ J. Thus, if w= x,Dis degenerate on{_{1, . . . ,}m} \J, whereas ifw= y,Dis degenerate onJ. In both cases, we obtain a contradiction.

The only case left is when D ⊆ {0, 1}^m admits a binary symmetric aggregator.

Contrary to the previous cases, where we showed that the respective non-dictatorial aggregators could not be generalized dictatorships, here we cannot argue this way, as Example13indicates. Interestingly enough, we show that as in Example13, we can always find some symmetric aggregator for such a domain that is not a general- ized dictatorship.

Claim 6. Suppose D ⊆ {0, 1}^m admits a binary non-dictatorial symmetric aggregator f = (f₁, . . . ,fm). Then, there is a binary symmetric aggregator g = (g₁, . . . ,gm)for D (g can be different from f ) that is not a generalized dictatorship for D.

Proof. If fis not a generalized dictatorship forD, we have nothing to prove. Suppose it is and let J ⊆ {1, . . . ,m}, such that f_j = ∨, for allj ∈ J and f_i = ∧for alli∈/ J(J can be both empty or{1, . . . ,m}).

LetD^∗ ={d^∗= (d^∗₁, . . . ,d^∗_m)|d= (d₁, . . . ,d_m)∈D}_{, where:}

d^∗_j =

(1−d_j ifj∈ J d_j else.

3.2. Generalized Dictatorships 47 By Lemma 4, h = (h₁, . . . ,h_m) is a symmetric aggregator for D if and only if h^∗ = (h₁^∗, . . . ,h^∗_m)is an aggregator for D^∗, whereh^∗_j = h_j, for allj ∈/ J and, for all j ∈ J, ifh_j = ∨, thenh^∗_j = ∧ and vice-versa. As expected, the property of being a generalized dictatorship carries on this transformation.

Claim 7. The operator h is a generalized dictatorship for D if and only if h^∗is a generalized dictatorship for D^∗.

Proof. Letx = (_{x1, . . . ,xm})_,y= (_{y1, . . . ,ym})∈ Dandz := _h(_x,y)_{. Since}∨(x_j,y_j) = 1− ∧(1−x_j, 1−y_j)and∧(x_j,y_j) =1− ∨(1−x_j, 1−y_j), it holds thatz_j = h^∗_j(x^∗_j,y^∗_j), for all j ∈/ J, and 1−z_j = h^∗_j(x^∗_j,y^∗_j), for allj ∈ J. Thus,z^∗ = h^∗(x^∗,y^∗). It follows thatz∈ {x,y}if and only ifz^∗∈ {x^∗,y^∗}_.

Now, sinceDadmits the generalized dictatorshipf, it follows thatD^∗admits the binary aggregator ¯∧ = (∧, . . . ,∧)

| {z }

m-times

, that is also a generalized dictatorship. Our aim is to show thatD^∗ admits a symmetric aggregator that is not a generalized dictator- ship. The result will then follow by Claim7.

For two elementsx^∗,y^∗ ∈ D^∗, we writex^∗ ≤y^∗if, for allj∈ {1, . . . ,m}such that x^∗_j =1, it holds thaty^∗_j = 1.

Claim 8. ≤is atotal orderingfor D^∗.

Proof. To obtain a contradiction, letx^∗,y^∗ ∈ D^∗ such that neitherx^∗ ≤ y^∗ nory^∗ ≤ x^∗. Thus, there existi,j ∈ {1, . . . ,m}, such thatx^∗_i = 1, y^∗_i = 0, x^∗_j = 0 andy^∗_j = 1.

Thus:

∧(x_i^∗,y^∗_i) =∧(x^∗_j,y^∗_j) =0.

Then, ¯∧(x^∗,y^∗)∈ {/ x^∗,y^∗}. Contradiction, since ¯∧is a generalized dictatorship.

Thus, we can write D^∗ = {d¹, . . . ,d^N}, whered^s ≤ d^t if and only ifs ≤ t. Let I ⊆ {1, . . . ,m}be such that, for all j ∈ I: d^s_j = 0 fors = 1, . . . ,N−1, andd^N_j = 1.

Observe that I cannot be empty, lest d^N = d^N−1 and that I 6= {1, . . . ,m}, since

|D| ≥3. Let nowg = (g₁, . . . ,gm)such thatg_j = ∧, for allj ∈ I andg_j = ∨, for all j∈/ I.

We show that g is an aggregator for D^∗. Indeed, let d^s,d^t ∈ D^∗ with s ≤ t ≤ N−1. Then, for allj∈/ I:

g_j(d^s_j,d^t_j) =∨(d^s_j,d^t_j) =d^t_j. Also for allj∈ I:

g_j(d^s_j,d^t_j) =∧(d^s_j,d^t_j) =0=d^t_j.

Thus, g(d^s,d^t) = d^t ∈ D^∗. Finally, consider g(d^s,d^N). Again, g_j(d^s_j,d^N_j ) =

∧(d^s_j,d^N_j ) = 0 for all j ∈ I and g_j(d^s_j,d^N_j ) = ∨(d^s_j,d^N_j ) = d^N_j , for all j ∈/ I. By definition of I, g(d^s,d^N) = d^N−1 ∈ D^∗. This, last point shows also that g is not a generalized dictatorship, since, for anys 6= N−1,d^N−1 ∈ {/ d^s,d^N}.

This completes the proof of Theorem13.

By Theorems10and13we obtain the following characterization.

Corollary 2(Dìaz et al. [9]). A domain D⊆ {0, 1}^m, with at least three elements, admits an aggregator that is not a generalized dictatorship if and only if there exists a possibility integrity constraint whose set of models equals D.

48 Chapter 3. Other forms of non-dictatorial aggregation- Characterizations of the corresponding domains

Remark 6. Apart from the initial issue regarding the non-degeneracy assumptions we dis- cussed in Remark4, here we have to deal with the additional requirement for the domain to be comprised of three or more elements. Thus, for this caseonly, we implicitly assume that we only consider possibility integrity constraints whose sets of models are comprised of at least three elements. In Section4.3, we argue that it is possible to distinguish such possibility integrity constraints from the rest (see Remark9).