[EM] Ultimate SPE Agenda Processing: Sink Swap Bubble

[Kristofer Munsterhjelm]

On 4/1/23 18:27, Kevin Venzke wrote:
> Hi Forest,
> 
> Le samedi 1 avril 2023 à 10:31:08 UTC−5, Forest Simmons <forest.simmons21 at gmail.com> a écrit :
>> Here's an example of another standard test case for Sink Swap Bubble ("Bubba" for short).
>>   
>> 48 C
>> 28 A>B
>> 24 B (sincere is B>A)
>>   
>> The smallest faction has thrown the sincere CW under the bus ... knowing that most Condorcet
>> methods, including classical wv methods like Ranked Pairs, would break the resulting ABCA
>> beat cycle at the weakest defeat A>B, leaving B as the winner.
> 
>> Bubble changes the order to B<A<C ... the finish order of the method ... thus disappointing
>> the defecting faction with a finish order polar opposite to their sincere preferences ...
>>   
>> When will they learn that you cannot mess with Bubba?
> 
> I've said this before, but I'm not a fan of this kind of method because, what if you're
> wrong about the sincere preferences? Then by electing C, you're actually punishing the A>B
> voters for not using compromise strategy. This muddies the water as to what behavior
> "chicken resistance" is actually incentivizing.

This is about chicken resistance in general, no? I suspect there's a 
more general impossibility result hiding here: that methods that are 
burial resistant and Condorcet (in the way the Smith-IRV hybrids or 
Friendly is) have to make uncharitable interpretations of ballots that, 
if they were honest, would imply a very different candidate should be 
elected. (I have no proof of this, but it seems a reasonable hunch.)

It's kind of like the "weak centrist! Condorcet winner! weak centrist! 
Condorcet winner!" thing cardinal proponents mention.

So if I'm right, then either we can choose a method that produces a very 
good result with honesty, or one that deters burial strategy, but not 
both. And then the ultimate deluxe method (in one sense) would be 
something along the lines of:

- Conduct an election with "honest" method X and "suspicious method" Y 
(or putting it differently, "compromise resistant" X and "burial 
resistant" Y).
- If the winners are equal, we're done.
- Otherwise do a manual runoff, where the voters' behavior will be 
honest (majority rule with two candidates is incentive compatible). 
Elect the winner.

But that might be asking for *too* much in the complexity department.

If you were to agree with the cardinal proponents, then X could even be 
something like Smith|Range (| being normalization) or STAR. But I tend 
to agree with rb-j that cardinal ballots are undefined. Even von 
Neumann-Morgenstern utility elicitation can be really difficult -- I've 
tried it a few times.

-km