Election Rigging
Nomenclature
• Let Xij be the (random) number of votes that will be cast for party i at station j, where the party index i takes
value [1,2] and the station index j goes over the range [1,…,N]
• Xj = X1j − X2j is a normal distribution with parameters {μj, σj}.
• This normal distribution is the compound probability distribution resulting from considering the multivariate
logistic-normal as the distribution for the probability parameter vector of the multinomial distribution. This
distribution is adequate to model multivariate count data when only the relative relationships between the
multinomial components are of interest.1
• There exist 2 models for modelling the winning event for the player 1 as is seen in the manner of conducting
elections in the States and in India.
• Model 1: W = 1A(ΣXj > 0), where j goes from 1 to n.
N
• Model 2: W = 1A(Σ1A(Xj > 0) > 2 ) , where j goes from 1 to n.
• zj : stu ed votes by party A (party index 1) in booth j.
• zĵ : upper limit on stu ng in station j, zj < zĵ
• Now there is a detection probability gi(zi) associated with station i which gives
the probability that Player 1 is caught for malpractice of stu ng zi votes.
• Lets assume this is convex increasing in zi.
• Threshold Model: 1 − (1 − g1(z1)(1 − g2(z2))) > α, where α is the upper
bound on the probability that Player 1 is caught for malpractice at either of the
stations.
ff
ffi
ffi
Threshold Model
• Theorem: (z1, z2) : 1 − (1 − g1(z1)(1 − g2(z2))) > θ is concave downward.
• No Inspector Case: The optimal value of (z1, z2) can be determined by nding the maximum
k, s.t, z1 + z2 = k touches the graph.
• One Inspector Case: Let z¯1and z¯2 be the maximum value of stu ng votes in booth 1 and 2
respectively, this is when the other booth has zero votes, for one inspector. The claim is that
( z¯1, z¯2) will be the best response with inspector being positioned at the argmax( z¯1, z¯2) station
number.
• Proof: If inspector deviates from choosing the argmax( z¯1, z¯2) station number while stu ng
party keeps same strategy, inspector stands to lose. If stu ng party departs from its strategy
while inspector remains in same position, it will only lose out since ( z¯1, z¯2) are already the
maximum values it can reach out to. ffi
ffi
fi
ffi
g
−1 1 1
(z ) − θ
z2 = g2 ( )
g1(z1) − 1
Concave Plot from 5 to 35 when
taking gi as the square function.
N
Model 2: W = 1A(Σ1A(Xj > 0) > )
2
• N=3 Booths, M=0 Inspectors Case Theorem: Given the initial probability of winning the
booths is p1 > p2 > p3 respectively and the probability of winning (2 or more seats) is given
as: W = p2 p3 + p1p3 + p1p2 − 2p1p2 p3
• Given su cient amount of stu able votes the end result will be either one of:
1
• p1 > p2 > p3 > with p1 taking on a large value
2
1
• p1 > > p 3 with p1 = p2 taking on large values in conjunction. • If the number of stu -able
2
votes is not su cient enough to cause p3 to become equal to the half, and the starting case
1 1
was > p1 > p2 > p3 it would remain > p1 > p2 > p3.
2 2
ffi
ffi
ff
ff
• We will look at the other setting where this δ is splittable, and we can change the pi at all
booths simultaneously. For this we will look at the problem as a nonlinear optimisation,
solve the KKT equations and get the optimal p* . This is less interesting because there is no
connection to zi, the actual stu ed votes in station i. Towards this, let pi(zi) denote the
probability of winning booth i when there are zi stu able votes. With abuse of notation,
pi(0) = pi . Let W(z̄) denote the probability of winning the election as a function of the
stu ed votes.
• Now pi(zi) = P(Xi > − zi) = Qi(−zi). For the rest of this section assume μi = 0 and σi = 1.
−zi2/2
• We will approximate pi(zi) using 1 − e . This follows from the cherno bound
−x 2/2
approximation for Q(x) ≤ e and using the fact Q(x) = 1 − Q(−x).
• Theorem: W(z) is maximised for z1 = z2 = z3 = Z/3, subject to Σzi = Z.
ff
ff
ff
ff
Non-Isomorphic Case
• In this section, the Qi function is taken in the general case to have μi = μi
and σi = σi .
Z − Σμi
• Theorem: W(z) is maximised for zi = max(μi + ,0) , subject to
3
Σzi = Z.
Convex Increasing Cost Function
• In this section, the constraint Σzi = Z is modi ed to introduce a convex
increasing cost of stu ng gi(zi). The constraint then becomes
Σgi(zi) ≤ G, where G is the total budget. Along with this a threshold is put
on the amount of stu ng that can be done on a speci c booth. This
constraint is given as zi ≤ z̄i
• A solution to this is proved with the water lling algorithm.
ffi
ffi
fi
fi
fi
N=1 Inspector Case in Model 2
• Amongst the M=3 booths, the position of the inspector is modelled by the
variable q1, q2 which are the probabilities of the inspector being in the
booth 1 and booth 2 respectively. The probability of the inspector being in
booth 3 is then: 1 − q1 − q2