A Borda Count For Partially Ordered Ballots

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 7

APATHY COUNTS: A BORDA COUNT PROCEDURE

FOR PARTIALLY-ORDERED BALLOTS

JOHN CULLINAN, SAMUEL K. HSIAO, DAVID POLETT

Abstract. The application of the theory of partially-ordered sets to voting systems is an important devel-
opment in the mathematical theory of elections. Many of the results in this area are on the comparative
properties between traditional elections with ranked ballots and those with partially-ordered ballots. In this
paper we present a combinatorial construction which allows for the generalization of the Borda count from
linearly-ordered ballots to arbitrary partially-ordered ballots.

1. Introduction
In this note we consider the problem of generalizing elections with linearly-ordered ballots to those with
partially-ordered ballots. In principle, partially-ordered ballots provide voters with greater flexibility for
expressing their true beliefs while still giving the option of submitting a traditional, linearly-ordered ballot.
The motivation behind introducing partially-ordered ballots is simple: it provides a platform that is less
restrictive than traditional ranked ballots. For example, given alternatives a1 , . . . , a6 , suppose a voter’s true
preferences are given by Figure 1, representing the fact that a1 is preferred to all alternatives; a2 is preferred

a1

a2 a5

a3 a4 a6

Figure 1. Partially-Ordered Ballot

to a3 and a4 ; a5 is preferred to a6 ; and there are no other preferences among the alternatives. The problem
is how to determine winners from the aggregate ballots.
Voting with partially-ordered preferences has been an active area of mathematics since Arrow’s seminal
work in 1951. For example, Brown (1974, 1975) explores much of the underlying set theory (filters) associated
with partial orders, acyclic relations, and lattice theory applied to voting theory. Ferejohn and Fishburn
(1979) introduce the notion of binary decision rules and apply it to a generalization of Arrow’s theorem;
similarly Barthélemy (1982) applies non-linear preference orders to generalizing Arrow’s theorem. More
recently, Fagin et al (2004) and Ackerman et al (2012) have studied special cases of partially-ordered ballots
– the so-called bucket orders – which are commonly referred to as weak orders or strict weak orders in the
social choice literature.
In Ackerman et al. (2012), the authors present a method they call bucket averaging that applies to the
special class of partially-ordered preference rankings they call bucket orders (these terms will be defined
below). The bucket averaging method is an approximation to the linear extension method. There, a
partially-ordered ballot is replaced by all possible linearly-ordered (traditional) ballots and the Borda count
is applied to these so-called linear extensions. The authors prove that their bucket averaging method gives
the same results as the linear extension method, and they further show that the computational complexity
of bucket averaging is less expensive than the linear extension method.
In this paper we describe a simple voting procedure for partially-ordered ballots that is inspired by the
classical Borda count for linearly-ordered ballots, and yields the classical Borda count results if all voters
submit linearly-ordered ballots. We place no restrictions on our partial orders (such as the bucket orders
above), and even allow for totally disconnected ballots.
We briefly describe our results. Detailed definitions appear in the next section. For a poset (P, 4) and
element x ∈ P , let a(x) be the number of elements in P to which x is preferred and b(x) be the number of
elements that are incomparable to x. Let Z≥0 denote the set of nonnegative integers. For each voter v and
poset (P, 4), we assign weights to the alternatives x ∈ P according to the formula w(x) = 2 a(x) + b(x). We
call this method of assigning weights the partial Borda score.
Theorem. The partial Borda score satisfies rigidity, deformation of Borda, and linearity. Furthermore, it
is the only scoring procedure that satisfies these conditions.
The theorem characterizes our voting procedure in terms of a few intuitively reasonable conditions (dis-
cussed in the next section). In the concluding section of the paper we will discuss further conditions that our
voting procedure does and does not satisfy. For instance, partial Borda count satisfies monotonicity but fails
to satisfy plurality. We also show that the partial Borda count specializes to the bucket averaging method
of Ackerman et al. (2012).

2. The partial Borda count


Our goal here is to develop a Borda count procedure for partially ordered ballots. We consider elections
where the voters submit ballots which consist of partially ordered preference rankings of the alternatives.
Such preference rankings can be represented by combinatorial objects called partially ordered sets (posets,
for short).
The terminology surrounding the theory of posets differs slightly among social-choice theorists and com-
binatorialists. We will follow the combinatorial conventions in Stanley’s book (1997). Let us recall basic
definitions and establish notation needed for the paper.
Definition 1. Let P be a finite non-empty set. Then 4 is a partial order on P , and P is called a poset, if
the following properties hold:
(1) Reflexivity: x 4 x for all x ∈ P .
(2) Antisymmetry: If x 4 y and y 4 x then x = y.
(3) Transitivity: If x 4 y and y 4 z then x 4 z.
Write x ≺ y if x 4 y and x 6= y. If x ≺ y and there is no z ∈ P such that x ≺ z ≺ y, then say that y covers
x. A poset can be visually represented by its Hasse diagram, in which elements of the poset are represented
by nodes, and a line is drawn from one node x up to another node y whenever y covers x. Several Hasse
diagrams are depicted in Figures 1 and 2. Two elements x and y in a poset are comparable if either x 4 y or
y 4 x. They are incomparable otherwise. A chain is a poset in which all pairs of elements are comparable.
An antichain is a poset in which all pairs of elements are incomparable.
Recently, Fagin et al. (2004) and Ackerman et al. (2012) have studied the aggregation of bucket-ordered
ballots. Our construction will apply to arbitrary posets, but for completeness we recall the definition of
bucket orders, which are a special kind of posets.
Definition 2. A poset (P, 4) is called a bucket order (or bucket poset) if P can be partitioned into a
disjoint union P = A1 ∪ A2 ∪ · · · ∪ Ak of nonempty sets (called buckets) such that for all x, y ∈ P , we have
x ≺ y if and only if x ∈ Ai and y ∈ Aj for some i, j such that i < j.
Figure 2 shows the Hasse diagrams of three bucket orders.
We now apply these combinatorial constructions to develop a Borda count procedure in which voters
submit ballots that consist of partially ordered preference rankings of the alternatives. Let us fix a set V
of voters throughout the paper. When presented with a set P of alternatives, each voter v must produce
a partial ordering of P (which could be graphically represented by a Hasse diagram) reflecting that voter’s
preferences. Such a poset constitutes a ballot. We will need to define a method for scoring alternatives on
individual ballots, as well as a method for determining social choices from the aggregate ballots.
Any method for scoring alternatives on individual ballots will be called a scoring procedure. Formally, a
scoring procedure takes as input a poset (P, 4), and gives as output a weight function w4 : P → Z≥0 , which
depends on the partial order.
2
Figure 2. Three Bucket Orders

Once a scoring procedure has been defined, we aggregate ballots according to the following procedure.
Fix the set of alternatives P and suppose that each voter v has submitted a ballot (P, 4v ). The score of
x ∈ P is defined to be X
s(x) = w4v (x)
v∈V
and the social choices are the elements of the subset SC of P defined by
(1) SC = {x ∈ P | s(x) = max{s(y)}}.
y∈P

It remains for us to define a weight function to be used in our scoring procedure.


For an arbitrary finite poset (P, 4) and x ∈ P , define the down-set A(x) and the incomparable-set B(x)
of x to be the following subsets of P :
def
A(x) = {y ∈ P | y ≺ x};
def
B(x) = {y ∈ P | y is incomparable to x}.
Let a(x) = |A(x)| and b(x) = |B(x)|. When it is important to emphasize the dependency on 4 we write
a4 (x) and b4 (x).
We propose the following method of assigning weights.
Definition 3. The partial Borda score is the scoring procedure that assigns a poset (P, 4) to the weight
function w4 : P → Z≥0 given by
w4 (x) = 2 a4 (x) + b4 (x).
The partial Borda count is the voting procedure whereby each voter v submits a poset (ballot) (P, 4v ) and
the social choices SC are determined according to Equation 1.
Example. Suppose that there are 6 voters and 5 alternatives, P = {a, b, c, d, e}, and the posets submitted
by the voters are as depicted in Figure 3. The vertices of each poset represent the alternatives. Next to each
vertex we indicate the weight assigned by the Borda scoring procedure. Note that the weights depend only
on the underlying poset and not on the voter who submitted it.

e 8 c 8
e 8
d 6 b 6
d 5 c 4 a 4
b 4 7d
c 3 b 6 b 2 d 2 5a e5 e8
a a c e a 0 a c
0 3 3 4d 4 e 0 2b c1 3 b3 3 d3

Figure 3. Partial Borda scores for several posets

The total weight of each ballot is 20 = 52 − 5. The scores of the alternatives are as follows: s(a) = 15,
s(b) = 23, s(c) = 22, s(d) = 27, s(e) = 33. Thus, the social choice is e.

3
We will show that the partial Borda score satisfies certain conditions which we believe to be reasonable,
and moreover partial Borda is the only scoring procedure satisfying such conditions. These conditions are
as follows:
(1) Rigidity: w4 (x) ∈ Z≥0 for
P all v ∈ V , all posets (P, 4) and x ∈ P .
(2) Deformation of Borda: w
x∈P 4 (x) = |P | 2
− |P | for all v ∈ V and posets (P, 4).
(3) Linearity: There exist r, s ∈ Z≥0 such that w4 (x) = r · a4 (x) + s · b4 (x) for all v ∈ V and posets
(P, 4) and x ∈ P .
Clearly linearity implies rigidity, so this list has redundancies. However, for the sake of clarity it is helpful
to see all the conditions stated explicitly.
√ The rigidity condition helps keep the voting procedure simple and the weights meaningful (a weight of
2, for example, would be difficult to interpret). The deformation of Borda condition ensures that the
total weight of the ballots is constant and agrees (up to a multiple of 2) with the classical Borda count if a
voter submits a linearly-ordered ballot (i.e., a chain). To explain further, recall that in classical Borda, each
voter ranks the alternatives in linear order x1 ≺v x2 ≺v · · · ≺v x|P | and then each alternative xi is given
P|P |
weight i − 1, so the sum of the weights is i=1 (i − 1) = (|P |2 − |P |)/2. In partial Borda, we would assign
weights w4v (xi ) = 2a(xi ) + b(xi ) = 2(i − 1). The scaling factor of 2 is necessary to keep the weights integral
for arbitrary posets. Finally, the linearity assumption says that, just as in the classical Borda count, the
weight attached to any alternative x by any voter v should depend only on a(x) and b(x) (of course in the
classical Borda count b(x) = 0) and this relationship should be linear. Thus, if a voter changes the preference
relation among elements of A(x) or B(x), but keeps a(x) and b(x) constant, then the weight attached to x
should not change. On the other hand, if a voter changes preferences by (for example) moving an element
from B(x) to A(x), then the weight the voter attaches to x increases by a constant amount, in this case r − s.

The partial Borda score clearly satisfies conditions (1) and (3). We now come to our main result, which
states that the partial Borda score also satisfies condition (2) and that these conditions uniquely characterize
the partial Borda score.

Theorem 1. The partial Borda score satisfies rigidity, deformation and Borda, and linearity. Furthermore,
it is the only scoring procedure that satisfies these conditions.

Proof. By definition the partial Borda score satisfies rigidity and linearity. We show by induction on |P |
that it also satisfies deformation of Borda. The case |P | = 1 is trivial. Suppose that partial Borda score
w4 (x) = 2a4 (x) + b4 (x) satisfies deformation of Borda for all posets of size n. Let (P, 4) be any poset of
size n + 1 and fix an element x0 ∈ P . Let (P 0 , 40 ) be the induced subposet obtained by deleting x0 from P .
This means for any x, y ∈ P 0 , we have x 40 y if and only if x 4 y.
For each x ∈ P such that x 6= x0 , we consider three cases:
• If x ≺ x0 then a4 (x) = a40 (x) and b4 (x) = b40 (x), so x≺x P0 w4 (x) = x≺x P0 w4 (x).
P P
0

• If x  x0 then a4 (x) = a40 (x) + 1 and b4 (x) = b40 (x), so xx0 w4 (x) = xx0 w40 (x) + 2 |{x ∈
P | x  x0 }|.
• If
P x is incomparable P to x0 , i.e., x ∈ B(x0 ), then a4 (x) = a40 (x) and b(x) = b40 (x) + 1, so
w
x∈B(x0 ) 4 (x) = x∈B(x0 ) w4 (x) + b4 (x0 ).
0

Putting all this together, we have


X X
w4 (x) = 2a(x0 ) + 2b(x0 ) + 2 |{x ∈ P | x  x0 }| + w40 (x) = 2n + n2 − n = |P |2 − |P |
x∈P x∈P 0

as desired.
Conversely, suppose we are given a scoring procedure with weight functions τ4 satisfying rigidity, de-
formation of Borda, and linearity. Thus τ4 has the form τ4 (x) = r · a4 (x) + s · b4 (x). By considering a
two-element chain x ≺ y, we have τ4 (y) = r, and by deformation of Borda we must have r = |P |2 − |P | = 2.
Similarly, by considering a two-element antichain we deduce that s = 1, and hence τ4 is the weight function
associated with the partial Borda score. 
4
Remarks. Elements x, y of a poset P are said to be equivalent if for all z ∈ P we have x ≺ z if and only if
y ≺ z; z ≺ x if and only if z ≺ y; and x and y are incomparable. It follows from our theorem that if x and
y are equivalent, then they are assigned the same weight. Thus, for bucket orders, all elements

3. Further Properties of the partial Borda count


This work was originally motivated by the results of Ackerman et al. (2012). There, the authors mainly
focus on aggregating ballots with bucket-ordered preferences. Given V and P , suppose (P, 4v ) is a bucket-
ordered poset for all v ∈ V . One way to determine the social choices from the (P, 4v ) is to replace (P, 4v )
with the (in principle, much larger) set of all linear extensions of 4v and then to perform a Borda count on
the linearly-ordered ballots. One of the main results of Ackerman et. al. is to show that the result of such a
Borda count is the same as their Bucket averaging method, which is defined as follows.
Let (P, 4) be a bucket order and partition P into subsets P1 , . . . , Pk according to 4 so that all elements
of Pi are equivalent and the elements of Pj cover those of Pi if and only if j = i + 1; set ni = |Pi |. The
bucket averaging method then assigns the following weight to each member of Pi :
[(n1 + · · · ni−1 − 1) + 1] + [(n1 + · · · ni−1 − 1) + 2] + · · · + [(n1 + · · · ni−1 − 1) + ni ] 2(n1 + · · · ni−1 ) + ni − 1
= .
ni 2

In other words, the alternatives in P1 each receive the average score of the first n1 Borda count scores,
and similarly for each Pi . We can immediately deduce the following result describing how partial Borda
score extends the bucket averaging scoring procedure.
Proposition 1. Suppose (P, 4) is a bucket poset. Then the bucket averaging method of Ackerman et al. as-
signs to each x ∈ P a weight of w4 (x)/2.
Bucket orders arise naturally in Borda count elections that allow for truncated ballots. In particular,
suppose we modify a traditional Borda count so that voters are allowed to rank a proper subset and give the
non-ranked alternatives a score of zero (an extreme version – so-called bullet voting – is when a voter only
ranks a single alternative). For example (using our convention of multiplying the Borda scores by 2), given
five alternatives a1 , . . . , a5 , a voter could give scores of 8 to a1 ; 6 to a2 , and 0 for a3 , a4 , a5 . A bullet vote in
this example would give a single alternative a score of 8, and all others a score of 0.
One can naturally create a bucket order from a truncated linear order by placing all unranked alternatives
into a single antichain. However, our partial Borda count would score these ballots differently than the
truncated Borda count, giving each of the unranked alternatives a score equal to the length of the antichain
−1, as opposed to a score of zero.
Proposition 2. The truncated Borda count and the partial count do not necessarily produce the same social
choices.
Proof. Suppose the following ballots are submitted in a truncated Borda count election, with the horizontal
lines indicating the truncation:
   
a b
b c
 ×5   × 4.
c d
d a
Then a is the social choice with 30 points, while b, c, and d receive 24, 16, and 8 points, respectively. However,
if we replace the five truncated ballots with

b c d

then overall a receives 30 points, while b, c, and d receive 34, 26, and 18 points, respectively, and b is the
social choice. 
5
Next, there are several properties related to monotonicity in the literature. A scoring procedure is said
to satisfy the monotonicity condition if x ≺ y implies x gets a smaller weight than y. The monotonicity
condition should be true of any voting system in which numerical data are attached to the alternatives and
the natural ordering of the integers is used to determine the social choices based on those data.
Proposition 3. The partial Borda score satisfies the monotonicity condition.
Proof. Let (P, 4) be a poset and suppose x, y ∈ P satisfy x ≺ y. Any element z ∈ P that is incomparable to x
is either incomparable to y or satisfies z ≺ y. Thus, w4 (y) ≥ 2 a(x) + (b(x) + 1) > 2 a(x) + b(x) = w4 (x). 

Remark. A function σ : P → Z≥0 that is monotone in the sense that x 4 y implies σ(x) ≤ σ(y) is
called a (reverse) P -partition. Thus, our weight functions w4 are examples of P -partitions. The theory of
P -partitions is a well developed area of combinatorics with far reaching applications. We refer the reader to
Chapter 4 of Stanley (1997) for further details and references.

Related to the idea of monotonicity is the consistency criterion, also called convexity when the underlying
geometry is emphasized (see Woodall (1994)): If the ballots are partitioned into two subsets and the social
choices of the resulting elections are the same, then that is the social choice of the original election. The
following is immediate from the definitions.
Proposition 4. Partial Borda count satisfies the consistency criterion.
A similar property that classic Borda and partial Borda share is that neither is clone-independent (see
Tideman (1987). This is well-known for classic Borda and the following is a counterexample for partial
Borda.
Proposition 5. Partial Borda count is not clone-independent.
Proof. Suppose |V | = 5, |P | = 6, and a clone a7 of a6 is inserted into a classic Borda count as depicted in
Figure 4. Originally, a1 is the social choice with 42 points and a6 is second with 40 points. In the cloned

a1 a6 a1 a6 a7
a2 a1 a2 a1
a3 a2 ×4 a3 a2 ×4
−→
a4 a3 a4 a3

a5 a4 a5 a4

a6 a5 a6 a7 a5

Figure 4. Clone Insertion

election, the social choices are a6 and a7 (with 45 points), while a1 receives 44 points. 

Lastly, we show that partial Borda count, unlike Borda count, does not satisfy the plurality condition: if
the number of ballots in which a is the single most preferred alternative is greater than the number of ballots
in which alternative b is shown any preference over another alternative, then a is preferred to b in the overall
election.
Proposition 6. Partial Borda count does not satisfy the plurality condition.
Proof. To see this, suppose there are 19 voters and three alternatives: a1 , a2 , and a3 . Partition the voters
into two subsets of size 10 and 9 with ballots as in Figure 5, respectively. Under the partial Borda count, a1
receives 40 points, a2 receives 28 points, and a3 receives 46 points. 
6
a3

a1 a2
×10 ×9
a2 a3 a1

Figure 5. Plurality Counterexample

References
[1] M. Ackerman, et. al., Elections with partially ordered preferences. To appear in Public Choice (2012)
[2] J.P. Barthélemy, Arrow’s theorem:unusual domain and extended codomain, Mathematical Social Sciences, 3, 79-89 (1982)
[3] G. Brightwall, P. Winkler, Counting linear extensions, Order 8, 225-242 (1991)
[4] J.D. Brown, An approximate solution to Arrow’s problem, Journal of Economic Theory, 9 (4), 375-383 (1974)
[5] J.D. Brown, Aggregation of preferences, Quarterly Journal Economics, 89 (3), 456-469 (1975)
[6] R. Fagin et. al., Comparing and aggregating rankings with ties. In: Proceedings of the 2004 ACM Symposium on Principles
of Database Systems, 47-58, (2004)
[7] J.A. Ferejohn, P.C. Fishburn, Representation of binary decision rules by generalized decisiveness structures, Journal of
Economic Theory, 21, 28-45 (1979)
[8] F. Roberts, B. Tesman, Applied Combinatorics, 2nd Ed. Pearson, New Jersey (2005)
[9] R. Stanley, Enumerative Combinatorics, Vol. 1, Cambridge Studies in Advanced Mathematics, vol. 49, Cambridge University
Press, Cambridge (1997)
[10] T.N. Tideman, Independence of clones as a criterion for voting rules, Social Choice and Welfare, 4, 185-206 (1987)
[11] D. Woodall, Properties of Preferential Election Rules, Voting Matters, 3, 8-15 (1994)

Department of Mathematics, Bard College, Annandale-On-Hudson, NY 12504


E-mail address: cullinan@bard.edu

Department of Mathematics, Bard College, Annandale-On-Hudson, NY 12504


E-mail address: hsiao@bard.edu

Department of Mathematics, Bard College, Annandale-On-Hudson, NY 12504


E-mail address: dp179@bard.edu

You might also like