Equal Shares and Then Some
Algorithmic Frontiers of Fairness
@FSTTCS 2023
Neeldhara Misra
Indian Institute of Technology, Gandhinagar
This talk is based on
...with additional inputs borrowed from the
wonderful website on the method of equal shares.
Neeldhara Misra
Indian Institute of Technology, Gandhinagar
Proportional Participatory Budgeting with Additive Utilities Dominik Peters , Grzegorz Pierczyński , and Piotr Skowron
The conference organizers want to build a registration kit.
₹ 1000 → Total Budget
300
270
350
700
50
50
300
270
350
700
50
50
300
700
300
700
A set of candidates \([m]\),
a set of voters \([n]\).
PB v. Multi-Winner Voting
Voters express preferences over candidates.
A subset of \(k\) candidates are chosen,
in a way that makes all voters happy.
Hopefully.
A set of candidates (projects)\([m]\),
a set of voters (stakeholders) \([n]\).
PB v. Multi-Winner Voting
Voters express preferences over candidates.
A subset of \(k\) projects are chosen,
in a way that makes all stakeholders happy.
This is the so-called unit-cost setting.
A set of candidates (projects)\([m]\) with associated costs,
a set of voters (stakeholders) \([n]\).
PB v. Multi-Winner Voting
Voters express preferences over candidates.
A subset of projects with a total budget of at most \(k\) are chosen,
in a way that makes all stakeholders happy.
If we are going to pick \(k\) winners,
a set of \(\frac{n}{k}\) voters who all
approve at least one candidate
should not be completely unrepresented, i.e,
at least one representative among the winners.
Justified representation (JR)
If we are going to pick \(k\) winners,
a set of \({\color{IndianRed}L} \cdot \frac{n}{k}\) voters who all
approve at least \({\color{IndianRed}L}\) candidates;
should not be unrepresented (in a stronger sense), i.e,
have at least one \({\color{IndianRed}L}\) representatives among the winners.
Extended Justified representation (EJR)
Given a ballot profile \(\mathbf{A}=\left(A_1, \ldots, A_n\right)\) over a set of candidates \(C\)
and a target committee size \(k\),
we say that a set of candidates \(W\) of size \(|W|=k\) provides justified representation for \((\mathbf{A}, k)\) if the following holds:
Justified representation (JR)
for every subset of voters \(N^* \subseteq N\) with \(\left|N^*\right| \geqslant \frac{n}{k}\),
either \(\bigcap_{i \in N^*} A_i = \emptyset\)
or \(A_i \cap W \neq \emptyset\) for some \(i \in N^*\).
Given a ballot profile \(\mathbf{A}=\left(A_1, \ldots, A_n\right)\) over a set of candidates \(C\)
and a target committee size \(k\),
we say that a set of candidates \(W\) of size \(|W|=k\) provides justified representation for \((\mathbf{A}, k)\) if the following holds:
Extended Justified representation (JR)
for every subset of voters \(N^* \subseteq N\) with \(\left|N^*\right| \geqslant \ell \cdot \frac{n}{k}\),
either \(|\bigcap_{i \in N^*} A_i| < \ell\)
or \(A_i \cap W \geqslant \ell\) for some \(i \in N^*\).
If the chosen candidates are \(W\), and a voter \(i \in [n]\) voted for \(A \subseteq C\), then her PAV score on W is:
\(s_i(W) := 1 + \frac{1}{2} + \cdots + \frac{1}{|A \cap W|}\).
Proportional Approval Voting (PAV)
The total PAV score of a committee is:
\(\sum_{i \in [n]} s_i(W)\)
PAV satisfies* EJR 🎉
If the chosen candidates are \(W\), and a voter \(i \in [n]\) voted for \(A \subseteq C\), then her PAV score on W is:
\(s_i(W) := 1 + \frac{1}{2} + \cdots + \frac{1}{|A \cap W|}\).
Proportional Approval Voting (PAV)
The total PAV score of a committee is:
\(\sum_{i \in [n]} s_i(W)\)
Determining a winning set for PAV is NP-complete,
even when every voter approves only two candidates.
If the chosen candidates are \(W\), and a voter \(i \in [n]\) voted for \(A \subseteq C\), then her PAV score on W is:
\(s_i(W) := 1 + \frac{1}{2} + \cdots + \frac{1}{|A \cap W|}\).
Proportional Approval Voting (PAV)
The total PAV score of a committee is:
\(\sum_{i \in [n]} s_i(W)\)
PAV's EJR guarantees
break (dramatically) beyond the unit-cost scenarios.
Proportional Approval Voting (PAV)
PAV's EJR guarantees
break (dramatically) beyond the unit-cost scenarios.
The Method of Equal Shares
100
100
100
100
100
100
100
100
100
100
Rule 1: We divide the budget equally among the voters.
The Method of Equal Shares
Rule 1: We divide the budget equally among the voters.
100
100
100
100
100
100
100
100
100
100
The Method of Equal Shares
Rule 2: We select the project with the highest number of votes.
100
100
100
100
100
100
100
100
100
100
300
270
350
700
50
50
6
5
5
7
4
5
The Method of Equal Shares
Rule 2: We select the project with the highest number of votes.
100
100
100
100
100
100
100
100
100
100
700
7
The Method of Equal Shares
Rule 3: The cost of the project is divided equally among its supporters.
100
100
100
100
100
100
100
700
7
100
100
100
The Method of Equal Shares
Rule 3: The cost of the project is divided equally among its supporters.
100
100
100
100
100
100
100
700
7
100
100
100
The Method of Equal Shares
100
100
100
100
100
100
100
100
100
100
300
270
350
50
50
6
5
5
4
5
The Method of Equal Shares
100
100
100
100
100
100
100
100
100
100
Rule 4: We do not count voters with no money.
300
270
350
50
50
6
5
5
4
5
300
270
350
700
50
50
300
270
350
700
50
50
The Method of Equal Shares
100
100
100
100
100
100
100
100
100
100
Rule 4: We do not count voters with no money.
300
270
350
50
50
0
3
3
1
2
The Method of Equal Shares
100
100
100
100
100
100
100
100
100
100
Rule 4: We do not count voters with no money.
300
270
350
50
50
0
3
3
1
2
The Method of Equal Shares
100
100
100
100
100
100
100
100
100
100
Rule 4: We do not count voters with no money.
300
350
50
50
0
3
1
2
The Method of Equal Shares
100
100
100
100
100
100
100
100
100
100
Rule 4: We do not count voters with no money.
300
350
50
50
0
3
1
2
Underutilized budget: 1000 - 700 - 270 = 30.
Equal Shares: The Method
Approval voting, where each voter votes for some of the projects, but votes for each of them with the same strength.
Utilities, where each voter assigns a utility number to each project
(0 or higher); each voter has the same number of total points.
The Input
or
Equal Shares: The Method
The overall budget is divided equally among the voters.
The Mechanism
If a project is not affordable* among voters who support it,
then delete the project from the system.
Equal Shares: The Method
The overall budget is divided equally among the voters.
We remove all projects from consideration that cost more than the combined share of all voters that voted for the project.
The Mechanism
Equal Shares: The Method
The overall budget is divided equally among the voters.
We remove all projects from consideration that cost more than the combined share of all voters that voted for the project.
If no projects remain, the computation of the method is finished.
The Mechanism
Equal Shares: The Method
The overall budget is divided equally among the voters.
We remove all projects from consideration that cost more than the combined share of all voters that voted for the project.
If projects remain, we calculate the effective vote count of every project.
The Mechanism
Equal Shares: The Method
The overall budget is divided equally among the voters.
We remove all projects from consideration that cost more than the combined share of all voters that voted for the project.
If projects remain, we calculate the effective vote count of every project.
Select the project with the highest effective vote count.
Split the cost of this project as equally as possible among supporters.
The Mechanism
Equal Shares: The Method
We do not count voters if they have already spent their entire budget share.
The Effective Vote Count
Voters who still have money left,
but not enough money to pay for the project when its cost is equally divided,
will count as a fraction.
Splitting Costs Equally
The cost of a project is split in such a way that
the maximum payment of any voter is as small as possible
among voters who approved it.
Equal Shares: The Method
0
0
30
30
10
60
0
0
1
1
0.4
Equal Shares: The Method
5
Equal Shares: The Method
10
Equal Shares: The Method
15
Equal Shares: The Method
20
Equal Shares: The Method
25
Equal Shares: The Method
30
Equal Shares: The Method
35
Equal Shares: The Method
40
Equal Shares: The Method
45
Equal Shares: The Method
50
Equal Shares: The Method
55
Equal Shares: The Method
60
Equal Shares: The Method
65
Equal Shares: The Method
70
Equal Shares: The Method
75
Equal Shares: The Method
80
Equal Shares: The Method
Equal Shares: The Method
>80
Equal Shares: The Benefits
Higher and more equal voter satisfaction.
Reduced bias towards the most popular categories.
Mathematical fairness guarantees.
ES v. other voting systems for participatory budgeting
Equal Shares: The Benefits
Support \(\implies\) Success
A project is guaranteed to be funded if
its exclusive support is proportional to its relative cost. 👍
Consider a project proposal with a cost that would consume 5% of the overall available budget.
Then, if at least 5% of voters vote for that project and for no other projects, then this proposal will win under the Method of Equal Shares.
Equal Shares: The Benefits
Theorem. Let \(P\) be a project and let \(\operatorname{cost}(P)\) be the cost of the proposal.
Suppose \(B\) is the overall budget, and \(n\) is the total number of voters.
If at least \(n \cdot {\color{IndianRed}\operatorname{cost}(P) / B}\) voters vote for \(P\) and no other projects,
then \(P\) will be among the projects selected by the Method of Equal Shares.
\(n \cdot \operatorname{cost}(P) / B\) voters
have a budget of at least \(B/n\) at the start.
\(\geqslant \underbrace{n \cdot \frac{\operatorname{cost}(P)}{B}}_{\text {number of voters }} \cdot \underbrace{\frac{B}{n}}_{\text {budget share }} = \operatorname{cost}(P) .\)
Their total leverage
Support \(\implies\) Success
Equal Shares: The Benefits
Theorem. Let \(P\) be a project and let \(\operatorname{cost}(P)\) be the cost of the proposal.
Suppose \(B\) is the overall budget, and \(n\) is the total number of voters.
If at least \(n \cdot {\color{IndianRed}\operatorname{cost}(P) / B}\) voters vote for \(P\) and no other projects,
then \(P\) will be among the projects selected by the Method of Equal Shares.
\(n \cdot \operatorname{cost}(P) / B\) voters
have a budget of at least \(B/n\) at the start.
\(\geqslant \underbrace{n \cdot \frac{\operatorname{cost}(P)}{B}}_{\text {number of voters }} \cdot \underbrace{\frac{B}{n}}_{\text {budget share }} = \operatorname{cost}(P) .\)
Their total leverage
Support \(\implies\) Success
Equal Shares: The Benefits
Groups with similar votes
Any group of voters who voted for similar projects can expect to be represented in the outcome
to an extent that is proportional to the group size.
A group of 20% of the voters can
expect to influence 20% of the budget spending.
Equal Shares: The Benefits
Groups with identical votes
Suppose that \(t\) out of the \(n\) voters submitted an identical ballot, that is, they all voted for the exact same set of projects.
Suppose: \(\operatorname{cost}\left(P_1\right)+\cdots+\operatorname{cost}\left(P_k\right) \leqslant \frac{t}{n} \cdot B,\).
In this case, the Method of Equal Shares will select all the projects.
Equal Shares: The Benefits
Groups with identical votes
Suppose that \(t\) out of the \(n\) voters submitted an identical ballot, that is, they all voted for the exact same set of projects.
Suppose: \(\operatorname{cost}\left(P_1\right)+\cdots+\operatorname{cost}\left(P_k\right) > \frac{t}{n} \cdot B,\).
In this case, the Method of Equal Shares will still be reasonable.
Equal Shares: The Benefits
Groups with identical votes
Equal Shares: The Benefits
Groups with identical votes
Theorem. Let \(T=\left\{P_1, P_2, \ldots, P_k\right\}\) be a set of projects, and
suppose \(t\) of the \(n\) voters voted for all of these projects
(and only these projects).
Then the Method of Equal Shares will select a subset \(T^{\prime} \subseteq T\) of projects such that for every project \(P^* \in T\) that was not selected by the Method, we have:
\(\sum_{P \in T^{\prime}} \operatorname{cost}(P)+\operatorname{cost}\left(P^*\right)>t / n \cdot B\).
Equal Shares: The Benefits
Groups of voters with overlapping votes
In practice, groups of voters with strong agreement
will not always vote for exactly the same set of projects.
In general, if \(T=\) \(\left\{P_1, P_2, \ldots, P_k\right\}\) is a set of projects, we write
\(\operatorname{cost}(T)=\operatorname{cost}\left(P_1\right)+\operatorname{cost}\left(P_2\right)+\) \(\ldots+\operatorname{cost}\left(P_k\right)\)
for the total cost of the projects in \(T\).
Equal Shares: The Benefits
Groups of voters with overlapping votes
In practice, groups of voters with strong agreement
will not always vote for exactly the same set of projects.
Suppose that a group of \(t\) voters all approve the projects in the set \(T\),
but some of the voters also approve some other projects not in \(T\).
Assume that \(\operatorname{cost}(T) \leqslant t / n \cdot B\).
Then, the Method of Equal Shares guarantees that at least one of the \(t\) voters will be sufficiently represented in the outcome \(W\).
Equal Shares: The Benefits
Groups of voters with overlapping votes
In practice, groups of voters with strong agreement
will not always vote for exactly the same set of projects.
Suppose that a group of \(t\) voters all approve the projects in the set \(T\),
but some of the voters also approve some other projects not in \(T\).
Assume that \(\operatorname{cost}(T) \leqslant t / n \cdot B\).
In other words, there exists some member \(i\) of the group with:
\(\operatorname{cost}\left(A_i \cap W\right) \geqslant \operatorname{cost}(T){\color{white}-\operatorname{cost}\left(P_j\right)} \text{ for some } P_j \in T\)
Equal Shares: The Benefits
Groups of voters with overlapping votes
In practice, groups of voters with strong agreement
will not always vote for exactly the same set of projects.
Suppose that a group of \(t\) voters all approve the projects in the set \(T\),
but some of the voters also approve some other projects not in \(T\).
Assume that \(\operatorname{cost}(T) \leqslant t / n \cdot B\).
In other words, there exists some member \(i\) of the group with:
\(\operatorname{cost}\left(A_i \cap W\right) \geqslant \operatorname{cost}(T){\color{IndianRed}-\operatorname{cost}\left(P_j\right)} \text{ for some } P_j \in T\)
Good-to-know Axioms
Core · FJR · EJR
Good-to-know Axioms
Core · FJR · EJR
Suppose that \(S\) can come up with a set \(T\) of projects
such that \(T\) can be funded with a \(|S|/|N|\) fraction of the budget.
Suppose further that each voter in \(S\) approves all the projects in \(T\); this means the group is cohesive.
EJR demands that the voting rule must select a set \(W\) such that at least one voter in \(S\) approves at least \(|T|\) of the funded projects in \(W\).
Good-to-know Axioms
Core · FJR · EJR
We say that a set \(S \subseteq N\) of voters blocks an outcome \(W\)
if there is a set \(T\) of projects affordable with a \(\frac{|S|}{|N|}\) fraction of the budget such that each member of \(S\) strictly prefers \(T\) to \(W\).
An outcome \(W\) is in the core if it is not blocked by any coalition \(S\).
EJR is a special case of the core where
the set \(T\) needs to be unanimously liked by the members of \(S\),
so that \(S\) is a “cohesive” group.
Good-to-know Axioms
Core · FJR · EJR
We say that a set \(S \subseteq N\) of voters blocks an outcome \(W\)
if there is a set \(T\) of projects affordable with a \(\frac{|S|}{|N|}\) fraction of the budget such that each member of \(S\) strictly prefers \(T\) to \(W\).
An outcome \(W\) is in the core if it is not blocked by any coalition \(S\).
For the approval-based case, it is unknown whether
there always exists an outcome in the core
(even under the unit cost assumption).
For general additive utilities, we already know that the core might be empty.
Good-to-know Axioms
Core · FJR · EJR
FJR strengthens EJR by weakening the cohesiveness requirement.
FJR requires that if a group \(S \subseteq N\) of voters can propose a set \(T\)
of projects that is affordable with \(S\)'s share of the budget,
and each voter has utility at least \(\ell\) for the set \(T\),
then at least one voter in \(S\) has utility at least \(\ell\) in the chosen outcome \(W\).
In the approval case, we see that FJR
does not insist that \(T\) is unanimously approved by the group \(S\) (like in EJR),
but it just requires that \(T\) is very popular among \(S\).
Good-to-know Axioms
Core · FJR · EJR
Both PAV and Equal Shares fail FJR.
That there does indeed exist a rule satisfying FJR,
which works for arbitrary costs.
(Not polynomial time, though.)
Equal Shares fails EJR, as does any efficient mechanism.
Equal Shares does satisfy EJR up to one project 😀
Equal Shares: The Benefits
Other Nice Things
Polynomial time computability
Independence of clones
(if a losing project is cloned; no clone wins,
if a winning project is cloned; at least one of the clones win.)
Voter Monotonicity
(more approvals on a winning project don't hurt the project)
Discount Monotonicity
(outcomes unchanged if a winning project becomes cheaper
or a losing project becomes more expensive)
Equal Shares: The Benefits
Other Not-So-Nice Things
Pareto-optimality
Budget limit monotonicity
(a winning project should stay winning if the available budget is increased)
Strategyproofness
(a voter should not be able to improve the outcome by changing their ballot)
Equal Shares: The Story
Equal Shares: The Story
Equal Shares: The Open Problem
The 74th constitutional amendment of India in 1992 brought with itself greater responsibility for the local government.
It defined 18 new tasks in the functional domain of urban local government such as slum improvement, urban planning and emphasized
greater citizen participation in local decision-making.
Equal Shares: Other Directions
A nice algorithm [sic] that satisfies FJR?
Axiomatic extensions to account for
scenarios involving groups and interactions?
Are structural domain restrictions useful,
as they turn out to be for multiwinner voting?
Does the core exist?
Fix the not-so-nice list for Equal Shares, especially PO?