rounding considered harmful - e-vote-id€¦ · dbms_random.normal random_nr from table...
TRANSCRIPT
Rounding Considered Harmful
Carsten Schurmann presented byPeter Rønne
IT University of CopenhagenDemTech
September 27, 2018
Carsten Schurmann (ITU, DemTech) Rounding Considered Harmful September 27, 2018 1 / 22
Article 76, Danish Election Law
(2) Each number of votes appearing as a result ofthe summation, shall be divided by 1 - 2 - 3 and soon until such number of divisions equivalent to themaximum number of seats expected to beallocated to the party or to the independentcandidate has been performed. The party or theindependent candidate having the highest resultingquotients shall be given the first seat in themulti-member constituency. The second highestquotient entails the second seat and so on and soforth, until all constituency seats in themulti-member constituency have been distributedamong the parties and the independent candidates.If two or more quotients are of equal size, lots shallbe drawn.
Carsten Schurmann (ITU, DemTech) Rounding Considered Harmful September 27, 2018 2 / 22
Inspecting the Implementation (in use)
select Obj_tillaegsmandater_landsdele
(Landsdel_id, Parti_id, Antal_stemmer,
Antal_kredsmandater, Kvotient_nr, Kvotient, Random_nr)
bulk collect into Temptab
from (select Landsdel_id,
Parti_id,
Antal_stemmer,
Antal_kredsmandater,
V_divisor Kvotient_nr,
round (Antal_stemmer / V_divisor) Kvotient,
dbms_random.Normal Random_nr
from table (P_col_tillaegsman_landsdele));
[Courtesy of Denmark Statistics]
Carsten Schurmann (ITU, DemTech) Rounding Considered Harmful September 27, 2018 3 / 22
Compensatory Seat Allocation, Parliament Election 2007
Carsten Schurmann (ITU, DemTech) Rounding Considered Harmful September 27, 2018 4 / 22
Observations
I The implementation of the Danish Seat Allocation algorithm rounds
I It looks like the error incurred by rounding is “neglibile”
But
I The quotients can get rather small (divisors > 30)
I The algorithm may erroneously draw lots
I Consequence: A wrong election result?
Carsten Schurmann (ITU, DemTech) Rounding Considered Harmful September 27, 2018 5 / 22
Summary of the talk
Do not round!
Carsten Schurmann (ITU, DemTech) Rounding Considered Harmful September 27, 2018 6 / 22
Overview of this Talk
1 Electoral Systems based on D’Hondt
2 Two Errors Due to Rounding
3 A Close Call
4 A Monte-Carlo Simulation
5 Conclusion
Carsten Schurmann (ITU, DemTech) Rounding Considered Harmful September 27, 2018 7 / 22
Electoral Systems based on D’Hondt
Carsten Schurmann (ITU, DemTech) Rounding Considered Harmful September 27, 2018 8 / 22
History
I Voting rule for Party-List Proportional Representation
I Seat Allocation Algorithm used in more than 50 countries
I Introduced by Thomas Jefferson for United States House of Representatives1791
I Introduced in Europe by Victor D’Hondt 1878
I Table computation using divisors 1, 2, 3, 4 . . .
I Estimates how many votes are necessary to win a seat
I Related Voting Rule: Sainte-Lague (different divisors 1,3,5,7 . . . )
Carsten Schurmann (ITU, DemTech) Rounding Considered Harmful September 27, 2018 9 / 22
D’Hondt and Drawing Lots
Without Rounding
A B C
Tallies 999 500 1501
Divisor 1 999 500 1501Divisor 2 499.5 250 750.5Divisor 3 333 166.6 500.3
Seat allocation (in order)
1. C,A,C,C
With Rounding
A B C
Tallies 999 500 1501
Divisor 1 999 500 1501Divisor 2 500 250 750Divisor 3 333 167 500
Seat allocation (in order)
1. C,A,C,A
2. C,A,C,B
3. C,A,C,C
Carsten Schurmann (ITU, DemTech) Rounding Considered Harmful September 27, 2018 10 / 22
D’Hondt and Drawing Lots
Without Rounding
A B C
Tallies 999 500 1501
Divisor 1 999 500 1501Divisor 2 499.5 250 750.5Divisor 3 333 166.6 500.3
Seat allocation (in order)
1. C,A,C,C
With Rounding
A B C
Tallies 999 500 1501
Divisor 1 999 500 1501Divisor 2 500 250 750Divisor 3 333 167 500
Seat allocation (in order)
1. C,A,C,A
2. C,A,C,B
3. C,A,C,C
Carsten Schurmann (ITU, DemTech) Rounding Considered Harmful September 27, 2018 10 / 22
Genuine and False Ties
Genuine Tie
A tie between two quotients is genuine if it is not caused by numeric effects, such asrounding.
False Tie
A tie between two quotients is false if it is caused by numeric effects, such asrounding.
Carsten Schurmann (ITU, DemTech) Rounding Considered Harmful September 27, 2018 11 / 22
Two Errors Due to Rounding
Carsten Schurmann (ITU, DemTech) Rounding Considered Harmful September 27, 2018 12 / 22
Error 1: Incorrect Allocation Order
Situation: More than k seats are to be allocated.For a k-way false tie the probability of allocating seats in the wrong order is
p =1
k!
Error 2: Incorrect Seat Allocation
Situation: Fewer than k seats are to be allocated.For a k-way false tie the probability of allocating seats correctly is
p =1(nk
)
Carsten Schurmann (ITU, DemTech) Rounding Considered Harmful September 27, 2018 13 / 22
Error 1: Incorrect Allocation Order
Situation: More than k seats are to be allocated.For a k-way false tie the probability of allocating seats in the wrong order is
p =1
k!
Error 2: Incorrect Seat Allocation
Situation: Fewer than k seats are to be allocated.For a k-way false tie the probability of allocating seats correctly is
p =1(nk
)
Carsten Schurmann (ITU, DemTech) Rounding Considered Harmful September 27, 2018 13 / 22
A Close Call
Carsten Schurmann (ITU, DemTech) Rounding Considered Harmful September 27, 2018 14 / 22
Error 1 Situation Observed, Parliament Election 2007
Carsten Schurmann (ITU, DemTech) Rounding Considered Harmful September 27, 2018 15 / 22
Hypothetical Error 2 Situation for Seats and 41, 42 (not in table)
Carsten Schurmann (ITU, DemTech) Rounding Considered Harmful September 27, 2018 16 / 22
A Monte-Carlo Simulation
Carsten Schurmann (ITU, DemTech) Rounding Considered Harmful September 27, 2018 17 / 22
Research Question
How likely is it that a Rounding Seat Allocation System commits to Error 1 and 2?
Carsten Schurmann (ITU, DemTech) Rounding Considered Harmful September 27, 2018 18 / 22
A Monto-Carlo Simulation for Error 1
Simulation parameters
I 8 parties
I Vote totals uniformly drawn from 20, 000− 400, 000
I 1, 000, 000 runs
Error 1: 9 out of 10 election results have 2-way false tie
n = 1 2 3 4
E (false n-way tie) 173.17545 0.904989 0.0048 0.000022
Carsten Schurmann (ITU, DemTech) Rounding Considered Harmful September 27, 2018 19 / 22
A Monto-Carlo Simulation for Error 2
Simulation parameters
I 8 parties
I Vote totals uniformly drawn from 20, 000− 400, 000
I 1, 000, 000 runs
Error 2: The probability of a false 2-way tie is roughly 1.5
n = 1 2 3 4
# observations 984, 089 15, 766 144 1
Carsten Schurmann (ITU, DemTech) Rounding Considered Harmful September 27, 2018 20 / 22
Conclusion
Carsten Schurmann (ITU, DemTech) Rounding Considered Harmful September 27, 2018 21 / 22
Conclusion
I Do not round!
I Recall that if b, d 6= 0 the following holds
a
b<
c
dif and only if a · d < c · b
Homework
Check if your country has implemented seat allocation correctly! Fix it, if not.Note: This homework only applies to delegates from d’Hondt and Sainte-Laguecountries
Carsten Schurmann (ITU, DemTech) Rounding Considered Harmful September 27, 2018 22 / 22