[EM] Election-Methods Digest, Vol 234, Issue 13

[Kristofer Munsterhjelm]

On 2024-01-12 19:45, Sass wrote:
>  > as of now I don't think anyone has much evidence for what will happen 
> in practice.
> 
> I think we do. We have the full ballot data on 448 RCV elections in the 
> US from this century. Only one did not have a Condorcet Winner. Even if 
> you reduce the set to elections with three competitive candidates 
> (defined as elections where the candidate with the third most first 
> choices has at least half as many as the candidate with the most), it's 
> still only 1 in 88, which could easily become 1 in 880 over time. If 
> elections with no Condorcet Winner are that unlikely, then by far the 
> strongest incentive for voters is to vote honestly as a rule. And we 
> know from RCV that voters are inclined to vote honestly under new 
> systems until the system backfires on them.

I think the problem is one of predicting how voters may alter their 
behavior when the circumstances change. Consider these possible 
descriptions:

- Voters always vote in a way that there's a majority candidate. If so, 
FPTP is sufficient.

- Voters always vote in a way that there's a number of no-hope fringe 
candidates as well as a mutual majority set containing clones of what 
would otherwise be a majority candidate. If so, IRV is sufficient.

- Voters always vote in a way that there's a Condorcet winner, possibly 
with spurious cycles from noise. If so, any Condorcet method will 
suffice, and Condorcet cycles can be handled like ordinary ties, by a 
coin toss or whatnot.

- Voters' honest distributions will always have a Condorcet winner but 
they may strategize, or be told to strategize by the candidates. If so, 
strategy resistance is more important.

- Voters will vote for multiple viable candidates if the method doesn't 
have too strong incentives to exit, and politics may evolve to be 
multidimensional, in which case honest cycles would appear. Then just 
how the Condorcet method deals with cycles would be important, as would 
robust clone independence (i.e. clone independence that generalizes to 
JGA's incentives to exit and entry).

- Voters have an absolute utility scale and would use it if they can, 
making distinctions beyond ranking. If so, we may need rated methods. 
(Or if a relative scale, something that normalizes rated ballots and treats

etc.

It's difficult to say ahead of time which of these are right. An 
argument to the extent that "we have n elections and none of these have 
shown behavior beyond the kth of these descriptions" has a flaw in that 
they are all under the context of the current method.

But we at least know that the first two descriptions are false. It *is* 
possible to say "ah, those two instances of center squeeze are just 
flukes" and keep going for IRV, but that seems rather iffy.

I suppose my position has been a combination of trying to get things 
right the first time (hence advanced/cloneproof Condorcet methods) and 
going by my own intuition (which finds the ambiguity of honest votes in 
a non-normalized rated system a real problem that burdens even honest 
voters with tactical decisions).

But I can't prove that "minmax and be there" would fail.

-km