[Election-Methods] Challenge Problem

fsimmons at pcc.edu fsimmons at pcc.edu
Sat May 31 11:53:23 PDT 2008


Dear Jobst, 

I have a another solution to the challenge problem:

p: A>C>>B
q: B>C>>A

Here's the method:

0. The ballots are approval with favorite indicated.

1. First the total approvals are counted and the candidates listed in order of approval.

2. For each ballot B, let f(B) be the approval score (as a percentage of the number of ballots cast) of the 
highest candidate on the list approved by B.

3. Let doc (for "degree of cooperation") be the average value of f(B) over all ballots.

4. Determine the winner with the lottery  RB*doc^M + RABMAC*(1-doc^M), i.e. decide whether to use 
Random Ballot or "Random Approval Ballot Most Approved Candidate" on the basis of the lottery
 (doc^M, 1-doc^M), where M will be specified below.

In the challenge problem, let's suppose that C is rated at R and S in the respective factions, with R > p  
and S >q, and that the approvals are

     x: AC
     x': A only
     y: BC
     y': B only

with x + x' = p, and y + y' = q.

Then as long as x + y > max(p, q) we have

doc = (x')^2 + (x+y)^2 + (y')^2, and

Prob(A wins) = doc^M*x'+ (1-doc^M)*p,
Prob(B wins) = doc^M*y' + (1-doc^M)*q,
Prob(C wins) = doc^M*(x+y).

For later use, note that doc=1 @ (x, y) = (p, q), 

and that the partial derivative of doc w.r.t. x or y @ (p, q) is 2,

so that the partial of doc^M w.r.t. x or y @ (p, q) is simply 2*M.

The "utility" expectation for the first faction is

E1 = P(A wins) + R*P(C wins)

If we evaluate the partial derivative w.r.t. x of E1  at the full cooperation point (x, y) = (p, q), 
we get

-1 - 2*M*p + R*(1 + 2*M) = 2*M*(R - p) - (1 - R)

which is greater than zero when M is sufficiently large, since R > p.
        

Similarly the partial derivative of E2 w.r.t. y at the same point is 

2*M*(S - q) - (1 - S),

which is greater than zero when M is sufficiently large.

Therefore, local unilateral defection from full cooperation won't pay if M is sufficiently large.

If I am not mistaken, the method is monotone and satisfies your property about proportional probability for 
those factions that steadfastly approve only their favorite.

On a technical note, if we replace the lottery  (doc^M, 1-doc^M) for deciding which kind of random ballot to use with 
the lottery (g(doc), 1 - g(doc)), where g(t)=1 - (1 - t)^(1/2), then we don't have to worry about M.

This works because (no matter how large M) the slope of g(t) eventually dominates the slope of  t^M as t approaches 
1.

Nevertheless, for the sake of simplicity I suggest using t^5 instead of g(t).

My Best,

Forest



 



More information about the Election-Methods mailing list