You are on page 1of 10

MICROECONOMICS OF

COLLECTIVE ACTIONS
A paper on sequential voting models

Сиволобов Максим Борисович, HSE, AE194


maxim.sivolobov@gmail.com
Contents
1. Intro. Discussion of Hemaspaandra article. Assumptions about typical online sequential game. ....... 2
2. Model .................................................................................................................................................... 2
Variables .................................................................................................................................................... 2
Election procedure .................................................................................................................................... 3
Herding ...................................................................................................................................................... 3
Voting strategies ....................................................................................................................................... 5
Unobservable order of voting ............................................................................................................... 5
Observable order of voting ................................................................................................................... 7
Does observing preceding votes in greater detail makes an average voter resistant to herding? .......... 7
Manipulation ............................................................................................................................................. 8
3. Conclusions, suggestions....................................................................................................................... 9
4. References ............................................................................................................................................. 9

1
1. Intro. Discussion of Hemaspaandra article. Assumptions about
typical online sequential game.

Many of online elections are conducted sequentially; the voter in them often knows, how many
people have voted for one of the variants before him, and what percentage of those that have
already voted. They constitute. In some cases, participant knows even the exact profile of
people, voting before him. In the majority of online voting systems participants do not know
exactly, how many more people will vote after them (there is, of course, some upper boundary,
like the total population of the social network or online community, but this typically an
unbinding restriction, as such voting typically voluntary, not compulsory) . Many elections in
real life are subject to manipulations, which may have disastrous consequences, both through
effects on current political situation and through destroying the overall credibility of democratic
institutions in society. Due to difficulty of prevention of manipulation proper, at some point, it
was suggested, that if one can not make eliminate the manipulation proper, one can, at the very
least, suggest an election system, which would be computationally resistant to manipulation, e.g.
while an optimal manipulation exists, it will take manipulators a prohibitively long (non-
polynomial) amount of time to find.
In their work, Hemaspaandra sisters and Roethe are discussing the manipulation in sequential
framework.
Of particular interest is the conclusion that, under natural election condition WCM and DVCM
are in P, e.g. can be solved in polynomial time by deterministic Turing machine. As the weighted
case is solvable in polynomial time, so is the unweighted. That has rather unfortunate
implications, meaning that
WCM-weighted coalition manipulation. DWCM-destruction weighted manipulation coalition.
In this work, we assume that besides the typical manipulation, available for simultaneous
elections, there is also a specific one, existing only in sequential elections and arising as a result
of herding – namely, manipulation of order. In their work , Ottaviani and Sorensen give an
example of a loss of information, that occurs, when too many exports give the same signal – the
consecutive experts will give the same signal, regardless of their private signal.
Depending on the amount of information the average voter possess, there can be several cases.
Our baseline case is when the voter knows, how many votes were given before him and their
distribution, but not the order, in which they were given and who have given them in particular.
They also are assumed to not know, how many votes were the people ready to throw for each
candidate after the public signal, but before the election procedure begins. They also do not
know who and in what quantity will vote after them (although there may be a reasonable upper
limit, it will be, in most cases, non-binding). The average voter also assumes that the other voters
are identical to him.

2. Model

Variables
C – a set of candidates; ci is the candidate preferred by i-th voter. There are n voters in total
V is a pair of voter’s name and his vote
2
pi is the vote of i-th voter in traditional election system, p’i – in median one
V<i is a ‘’snapshot’’ of elections (similar to Hemnaspaarda definition)
qij is probability, assigned by the i-th voter to an occasion, that j-th candidate is the best
candidate; Qi is the combination of such probabilities for an i-th voter.
α is a herding parameter (assumed to be identical across voters for simplicity)
l is the initial public signal, which can be not attributed to any of the voters in particular
nji – number of votes given to the j-th candidate before i-th voter votes
ni – number of ‘’votes’’ given for the preferred by i-th voter before i-th voter votes himself.
ne(i) – number of ‘’votes’’, given to the candidate(s), which would have been elected, if the last
voter was the voter before the i-th one

Election procedure
Timeline:
1) Initial opinions of the voters about the candidates are formed. These opinions are perceived
probabilities that the particular candidate is the best one
2) A public signal, observed by everyone, happens, which results in public re-arranging the pre-
voting expectations. Public signal is assumed to be considered at least somewhat informative by
all voters.
3)cSequential voting takes place. For every voter but the first one, the opinions about the
candidates are affected by publicly observed choices of all preceding voters
4) ‘Votes are counted, if no tie – victor is determined, otherwise -randomization between those
who tied
Each voter selects a candidate he votes for. His choice may be his most preferred candidate or
not. The actual winner of the elections is the candidate with the median number of votes (or the
nearest to the median from above, if the number of candidates is even). The position, in which
candidates are elected, or are tied for election, if voting ends now, is further called the elected
position.
Each voter has own opinion about candidates’ quality, which is represented by qij, Sum of all qij
is equal to 1. Voter chooses the candidate, whom he perceives to have the largest probability to
be a good one, after combining his own opinion with inferences about the quality from public’s
opinion of those voters, that precede him in voting.
Voters are assumed to know, how many people have voted before them and how many will vote
after.

Herding

Initially, politicians through public signals influence the opinion of society about them. Each
voter formulates the opinion about each candidate – namely, the probability, given th public
signals, that that candidate is the best. After the election starts, decisions of preceding voters also
3
become public signals of sorts. Voters have some perceived average quality assigned to signals
of preceding voters. The reasons for assigning positive probability to others’ signals may vary.
For instance, there might be a knowledge, that amongst the voters there are some experts with
precision of signal above 0,5, while the As long as the perceived average quality of public voter
signal is not equal to 0,5, e.g. it is considered informative, there exists propensity for herding (or
anti-herding).
Our model was inspired by the model of Ottaviani and Sorensen, used in their work ”Information
aggregation in debate: who should speak first?’’. In particular, there are more alternatives, than
2, voters treat each other as a homogenous set of relatively weak exports (for lack of better
information). They all receive the same public signals, but formulate different priors. A voter, in
essence, wants to know, which candidate the preceding voter(s) would choose, if they had access
only to information dispersed through the initial public signal. Instead, he gets information, what
candidate the preceding voters will choose, given his own original public signal interpretation
plus the signals produced by voters before him.
In the base case, we assume that the average voter does know the distribution of votes taken
before him, but not the preferences of people who have taken them. Knowledge of how many
people will vote after him, is non-essential for the non-manipulating voter in the traditional
scheme, where the most voted for candidate wins. He assumes, that the voters with various
preferences are distributed amongst the voting queue randomly. Based on preceding votes, he
may try to infer, which candidate is more popular by Bayesian updating. His own weight takes
into account his own initial opinion about the candidates and his impression of the public opinion
about, constructed based on preceding votes. If voters significantly care about public opinion,
instances of herding may occur. A case may arise, when one of the candidates initially (by
manipulation or by pure accident) had a notable majority of votes, which leads to herding among
consequent voters, which may be hard to reverse... In presence of herding, sequential voting
system, where one can observe preceding votes, is unstable. While, assuming no manipulation,
the most popular candidate before the start of sequential voting is the one most likely to win,
such voting, in presence of herding, has an undesirable property of probably being determined by
the relatively small portion of initial voters; simultaneous unmanipulated elections do not have
that problem.
The update of the perceived quality of candidates in the traditional voting scheme goes like this:
𝑞𝑖𝑗 (𝑖 − 1) ∗ 𝛼
𝑞𝑖𝑗 = 𝐼(𝑐𝑖−1 = 𝑗) ∗ + (1 − 𝐼(𝑐𝑖−1 = 𝑗))
𝑞𝑖𝑗 (𝑖 − 1) ∗ 𝛼 + (1 − 𝑞𝑖𝑗 (𝑖 − 1)) ∗ (1 − 𝛼)
𝑞𝑖𝑗 (𝑖|𝑖 − 1) ∗ (1 − 𝛼)
∗ , 𝑤ℎ𝑒𝑟𝑒 𝑞𝑖𝑗 (𝑖 − 1)
𝑞𝑖𝑗 (𝑖 − 1) ∗ (1 − 𝛼) + (1 − 𝑞𝑖𝑗 (𝑖 − 1)) ∗ 𝛼
= 𝑞𝑖𝑗 |𝑙, 𝑝1 , … , 𝑝𝑖−2

𝑞𝑖𝑗 (𝑖 − 2) ∗ 𝛼
𝑞𝑖𝑗 (𝑖 − 1) = 𝐼(𝑐𝑖−2 = 𝑗) ∗ + (1 − 𝐼(𝑐𝑖−2 = 𝑗))
𝑞𝑖𝑗 (𝑖 − 1) ∗ 𝛼 + (1 − 𝑞𝑖𝑗 (𝑖 − 1)) ∗ (1 − 𝛼)
𝑞𝑖𝑗 (𝑖|𝑖 − 2) ∗ (1 − 𝛼)
∗ , 𝑞𝑖𝑗 (1) = 𝑞𝑖𝑗 |𝑙
𝑞𝑖𝑗 (𝑖 − 2) ∗ (1 − 𝛼) + (1 − 𝑞𝑖𝑗 (𝑖 − 2)) ∗ 𝛼

4
If we switch to median winner voting system, thar should, intuitively make manipulations with
order less effective. As it is now less obvious, how many more people exactly have supported the
guy in the median spot. While the perceived individual signal strength of preceding voters does
not change, the strength of their combined signal deafens, e.g. becomes closer to 0.5. The
expected number of people, who have (so far) supported a particular candidate in the election,
probably can be estimated, although, given many possible combinations of who orders of voting,
not particularly fast, though in polynomial time; the original number of people amongst those
voted, who were going to vote for a particular candidate, in many cases, likely cannot be
estimated. The individual impact of a singular vote on the opinion of other voters, assuming
herding, is weakly weaker in median sequential voting system, than in traditional voting system.
Below, one can see a mechanism of perceived quality update for the case when voters can
observe all voting order before them
𝑞𝑖𝑗 ′(𝑖 − 1) ∗ 𝛼
𝑞𝑖𝑗 ′ = 𝜋(𝑖 − 1) ∗ + (1 − 𝜋(𝑖 − 1)
𝑞𝑖𝑗 ′(𝑖 − 1) ∗ 𝛼 + (1 − 𝑞𝑖𝑗 ′(𝑖 − 1)) ∗ (1 − 𝛼)
𝑞𝑖𝑗 (𝑖|𝑖 − 1) ∗ (1 − 𝛼)

𝑞𝑖𝑗 ′(𝑖 − 1) ∗ (1 − 𝛼) + (1 − 𝑞𝑖𝑗 ′(𝑖 − 1)) ∗ 𝛼

, 𝜋(𝑖 − 1) = 𝐸𝑃(𝑐𝑖−1 = 𝑗|𝑝1 , … , 𝑝𝑖−1 ), 𝜋(1) = 𝐸𝑃(𝑐1 = 𝑗|𝑝1 ), 1 ≥ 𝜋(𝑖 − 1) ≥ 0

In the unobserved case, since voters cannot observe 𝑝𝑖 -s directly (unless they are the second
voter), they estimate total amount of people who supported each candidate at the moment of vote
based on nji-s and from that estimate the number of signals given by the voters for and against a
particular candidate.
It should be noted that a signal, given by voter against somebody else, can be treated as a signal
in favor of everybody else.
Sequential voting is different from simultaneous voting in that, if there is some propensity of
population to heard, there is a whole new way of manipulation – a manipulation of order. It has
nothing to do with changing the voter’s vote directly, e.g. by bribe – it is simply about the
rearrangement of order of votes , such that the ones initially supporting the manipulation
coalition’s candidate are grouped in the beginning of the voting queue, followed by those with
large propensity to herd. If the voters of one of the candidates are more disciplined or
determined, then the voters of others, they can show en masse early, severely skewing the
elections in favor of their preferred candidate.

Voting strategies
Unobservable order of voting
How do the voters vote in sequential voting, For standard case, this is straightforward – in
absence of manipulation, they simply vote for their preferred candidate. The voting in median-
winner system, however, depends on the position, in particular , whether the preferred candidate
is in the elected one already. Voters in the baseline case will try to make the preferred candidate
closer to being the median one. If candidate is above the elected, they vote for one of the
currently elected ones. If only one candidate is in the elected position and that vote doesn’t make
him share position with somebody else, this move will be supported by the next voter, supporting

5
candidate above the elected position. Due to the nature of the game, there always exists a vote
that improves or at least not worsens the current position of the preferred candidate- that is, not
diminishes his chances of getting elected compared to the case, in which i-th voter did not make
a vote – the voters, supporting the candidate not in the elected position, always have the option
of either moving their preferred candidate closer to the elected position, or moving some
candidates in the elected position closer to them: supporters of the candidate in the elected
position can always move some other candidate farther from it. When the preferred candidate is
above the elected position (ni>ne), the one preferring him is interested in upvoting those from the
median one to the closest below – and knows, that those others above the median have the same
interests, and position of everyone above median is improved if the voter above median is
upvoted. Everyone who favors the candidates above the one in the elected one will be upvoting
the median one. As can be noted, the strategy of supporters of the candidates below the elected
position What will the supporters of the one currently being the median, do? If there are ties,
they first check, whether by upvoting their candidate they can get him into an election position
with fewer number of ties, then before; if not and there are ties, they upvote one of the tied
contesters; if they are no tied contestants, they either upvote their candidate, or upvote the closest
candidate to the election. They can either upvote their candidate, to increase the distance from
those pursuing from below, but this decreases distance to those above the median, or they can
upvote the nearest candidate above theirs. The candidates above the median present more danger
in terms of catching the current unique leader; however, being ousted from the elected position
into position above it is not as bad, as being ousted into position below it; therefore, if , looking
at current proportion of voters and trying to estimate the expected proportion, supporting each
candidate, assuming herding is a random occurrence, assumes that the candidate above him is
more likely to catch him, but the amount of trouble it causes for him is less than being caught in
the by the candidate below, might change his behavior as the expected number of voters left
increases. While candidates above the leading one are usually expected to have more supporters,
than the candidates below (in each of those two respective subgroups, the closer to the median,
the higher is the expected level of support; the current leader in unobservable case, always has
the highest expected number of people who actually supported him at the moment of their vote),
sometimes, where there are few voters left before the end, there might be positions, where
pursuers from below might reach leader, while those from below can’t upvote him enough to tie
him for the elected position.
Strategies for voters for voters are subdivided into two cases: 1) strategies of those, who support
voters candidates not in the elected position, which are trivial and do not change with expected
(or actual) number of voters left after them. 2) strategies of supporters of the candidates in the
elected position, which depend not only on the fact that their preferred candidate is in the elected
position, but also, if the candidate is in the elected position alone, on how close are the pursuers
from below and from above.
For those voters, who support the candidates below and above the elected position, once they
have decided upon their preferences, the expected proportions on of the voters for their
opponents matter no longer. It is not the case for the supporters of the candidate(s) currently in
the unique elected position. If the preferred candidate is currently in the unique elected position,
he has to choose, whether to increase distance from those below him or from those above him.
While in unobserved order case, those above the elected position are usually perceived to have
more fewer people, who have voted while supporting
Therefore, we can outline the sources of ambiguity of who voted for whom. An upvote for the
candidate below the elected position, in absence of manipulation, as well as an upvote of a
6
candidate above the elected position, can be to the benefit of, respectively, supporters of the
upvoted candidate below the elected position and (some of the) median candidate(s). The upvote
of the candidate in the elected position can be, in certain situations, attributed to either his
supporters or supporters of candidates above the elected position.
Observable order of voting
The crucial difference between observable order of voting and unobservable order of voting lies
in the fact, that, in median-winner system, if the exact order of votes can not be observed, the
candidate(s) currently at the elected position are perceived to have the most largest number of
voters so far, who supported them, while voting. This trait is shared with a traditional voting
system – the one currently in the leading position is known to have the largest proportion of
supporters out of those who voted so far, assuming no manipulations. In observable median-
winner games that is not necessarily the case – the candidate, who managed to stay in a unique
elected position for many consecutive periods may have higher perceived amount of supporters
than the one that got into unique elected position recently. That means, that the opponent, that is
expected to be the toughest for the candidate not currently in the elected position, is not
necessarily a candidate or candidates, currently in the elected position

In traditional game, assuming you observe all votes and no manipulation happens, you can
always clearly connect the vote and the candidate the voter supported at the moment the vote
was submitted. In the median winner game, even if you observe all the votes as they come, there
can be situations, when there one cannot tell clearly, whether the voter was a supporter or an
opponent of a particular candidate (and, if he was an opponent of this particular candidate, which
of the other candidate he supported?).

Does observing preceding votes in greater detail makes an average voter resistant to
herding?
Other variant of the elections is when the candidates know all the preceding votes and the order,
in which they were conducted, but still do not know who voted for what, and many cases do not
know precisely, whose supporters has cast their particular vote. In this case, the voter should
somehow correct his signal for herding. He does not know pre-voting beliefs of other consumers,
and he also does not know the level of precision which the voters before him attribute to signals
of voters before them. When one can not observe the order, in which the election proceeds, one
can assume, that herding in the preceding voters behavior happens randomly, and thus views the
votes people give in to be, on average, representative of their own initial opinion. In this case, a
long preceding chain of voters, who vote the for the same candidate, can make one suspicious,
that herding may have taken place and votes taken do not represent the initial opinion of some
voters. Thus, one should somehow correct for apparent herding. This correction is likely not
possible, unless the voter have some assumptions about the initial preferences of other voters.
What if, after observing the initial signal of the politicians, the voter also formulates some
predictions about what portion of the voters are expected to support each candidate, or he has
some external estimates (like pre-election polls). Observation of votes of preceding voters not
only changes the opinion about the candidates, it also changes, in the eyes of the voter, the
expected proportion of population, supporting them.
Observation of separate preceding votes, in case of median winner scheme, may look like it
allows for more precise estimation of number of supporters (so far) of each candidate. If the
voter knows the average preferences of voters, initially supporting candidates other than his own,
7
and their average trust of the public signal, he may attempt to estimate the probability that some
of the preceding voter herds, and thus his opinion should not be taken into account.

𝑞𝑖𝑗 |𝑙, 𝑝1′ , … , 𝑝𝑖−1 = 𝑞𝑖𝑗 |𝑙
𝑖−1
∗∏ 𝐼(𝑝𝑘
𝑘=1
𝛼
= 𝑗) ∗ ( ∗ (1 − ℎ(𝑖|𝐼 = 1) + 1 ∗ ℎ(𝑖|𝐼 = 1)) + (1 − 𝐼(𝑝𝑘 = 𝑗)) ∗ ((1 − 𝑎)
1−𝛼
∗ (1 − ℎ(𝑖|𝐼 = 0)) + 1 ∗ ℎ(𝑖|𝐼 = 0))

Where h(i) is the perceived probability that the i-th person is herding. This formula works only,
if we know average propensities to heard among the groups and have some pre-signal
information about the voter’s preferences. The problem is, even if we do know the voters’
distribution before signal (nothing says that this exit poll cannot be manipulated), voters’ average
propensity to herd, we cannot estimate, how we should correct for herding if we don’t know the
public signal, obtained by probably herding voter. One cannot estimate the probability of voter
herding, given the votes preceding the voter’s, if one does not know, against how good of an
alternative the voter herds. In other words, while trying to correct for herding with purpose of
getting information of the preceding voters’ public signals, one can only do it, if one knows
those signals in the first place, which invalidates the correction procedure.
So, in essence, observing the order of preceding votes gives no particular advantage to the
average voter (aside from knowing the vote of the first guy, not influenced by other voters,
which is not a huge consolation) in terms of resistance to herding. We assume, that the
manipulator actually knows, what signals that kind of person receives – otherwise, he may have
issues figuring out the most optimal way of manipulation. While absence of knowledge about the
personal signal obtained does not necessarily invalidate manipulation, it makes the manipulation
coalition weakly worse at it.

Manipulation
For the purposes of this section, the manipulator(s) are assumed to know both the order, in which
the voting happens, the pre-signal preferences of the voters and how they change with response
to signals – basically, everything. If we are in a traditional sequential game with herding, the
manipulations affect the choices of non-manipulating voters through one channel – namely,
through the change of the opinion of voters about candidates. In the case of median-winner
system, there are 2 additional channels:1) the candidate may be moved by votes of the
manipulation coalition into a position, which will force his supporters to change strategy 2)
while the position of the candidate and the beliefs of his voters may not be changed, their
perception of the optimal strategy may be affected (for voters of the candidate in the unique
elected position).
Manipulators, willing to optimize their manipulation, should take into account not only the
change of the preferred candidate for the voter they are swaying and the position of the voter
relative to the other voters, but also the comparative standings of the candidates relative to the
election position.
In a traditional game, lots of relatively weak opposing candidates were beneficial for the leader,
who aimed to coerce voters to vote for his preferred candidate through the process of herding –
lots of weak opponents meant that the chain of voting for one particular candidate (not the one
preferred by manipulator), was relatively unlikely to appear. Situation in a median-winner game
8
with unobservable order is not quite so clear-cut. Assuming that the initial proportion of the
supporters, going to vote for the candidate, preferred by the manipulators stays the same, higher
m means, that, provided that candidate, preferred by manipulators, gets in the unique elected
position, more votes are expected to separate him and the candidates below the elected position,
as those candidates are individually weaker. The individual weakness of the voters above the
election position is not so much of an advantage, as, given that the overall proportion of total
supporters of candidates above the election position hasn’t changed, the candidate in a unique
leader position is no less likely to be upvoted by them, while his supporters will probably have to
spread their upvotes between various candidates above the leader, resulting in smaller probability
of preferred candidate retaining leadership. In an unobservable order median-winner game an
increase in number of candidates will increase the threat from those above the median and
decrease a threat of those from below, given that the number of people, supporting the leader,
stays the same.

3. Conclusions, suggestions

Due to the fact, that public signal generated by voter, is weakly weaker in median-winner
system, coercing voters through invoking herding may be somewhat harder in terms of amount
of resources spent. For the leader, establishing a significant lead over some of the pursuers from
above is expected somewhat harder in an unobservable order median-winner elections, than from
pursuers in highest-voted-wins traditional elections.
Analysis of computational difficulty of median-voter system required. Further investigation in
behavior of various categories of voters in the observable case required.

4. References

1. P.Faliszewski, E.Elkind, P.Bachrach “Coalitional Voting Manipulation: A Game-Theoretic


Perspective” (2011)
2. P.Faliszewski, A.D.Proccacia “AI War On Manipulation: Are We Winning?”(2010)
3. E.Hemaspaandra, L.A. Hemaspaandra, J.Roethe ‘’The Complexity Of Online
Manipulations Of Sequential Elections’’ (2013)
4. M,Ottaviani, P.Sorensen “Information aggregation in debate: Who Should Speak First?”
(2000)
5. P. Faliszewski, E. Hemaspaandra, and L. Hemaspaandra. ‘’How hard is bribery in
elections?’’(2009)

You might also like