[EM] IRV proponents figure out how to make IRV precinct-summable

[Kristofer Munsterhjelm]

Dave Ketchum wrote:
> 
> On Mar 17, 2009, at 7:09 PM, Kristofer Munsterhjelm wrote:
> 
>> Kathy Dopp wrote:
>>> On Tue, Mar 17, 2009 at 1:19 PM, Dave Ketchum 
>>> <mail.clarityconnect.com> wrote:
>>>> There has been a lot of guessing - let's see if I can do better, though
>>>> wishing to move to Condorcet:
>>>>
>>>> Precinct-summable IRV is not reachable.  The first counts of top 
>>>> ranks have
>>>> to be centrally summed to identify certain losers.  Then for each 
>>>> ballot of
>>>> such a loser the next-ranked not-yet-lost candidate must be reported.
>>>> Choices here are:
>>>>     Have precinct do it, since they have the ballots.
>>>>     Have had ballot images forwarded so central can do the count.
>>>>
>>> I agree with you David. And BTW, I also agree with most of what
>>> Kristofer said and in retrospect did not mean that his software idea
>>> was "vaporware" so much as that it would take at least 4 years to be
>>> federally certified at the cost of hundreds of thousands of dollars
>>> just for the certification so that it is "vaporware" only from that
>>> sense of not being available for most states to use for a long time.
>>
>> I would say that the only way to make it summable is to do it my way, 
>> or at least emulate my way. From what you say, it seems that they make 
>> it "summable" by eliminating all but two candidates and then seeing 
>> which one wins; that is, they run a "fake first round" for all 
>> possible combinations of winner candidates. Then a Plurality count 
>> determines who those two winner candidates are. My claim is then that 
>> of some winner set {X, Y} of two candidates (possible two round 
>> result), X wins iff c[X, Y] > c[Y, X]. That means that their method is 
>> a hackish variant of mine, where the hack is required because they're 
>> stuck with currently certified voting machines.
> 
> Summability is a feature of an election method.  I suspect that what you 
> offer fails to pick the winner IRV would, and/or takes too much effort 
> to deserve bragging.
> 
> Condorcet is summable because ALL of the information from a ballot can 
> be copied into the N*N matrix on first reading of the ballot.
> 
> In IRV at the time that the A of A>? is read what the "?" may be does 
> not matter.  When it is determined that A is a loser, then the "?"s on 
> those ballots will matter.

It fails to pick the winner IRV would, but it picks the winner the 
"contingent vote" summability hack picks. The contingent vote is like 
this: first do a plurality count. The two candidates that have the 
greatest count go to the second round, where the one that's ranked above 
the other more often wins. The second round is thus a pairwise 
comparison, but it's not Condorcet, since it doesn't check all pairwise 
comparisons.

>  From this distance I do not know whether what was done in Cary was valid:
>      If initial counts were 49A, 48B, and 3other, A or B will win and 
> all the others can be disposed of together.
>      If initial counts were 27A, 26B, 25C, and 22D, D loses.  Suppose 
> counts then are 35A, 34B, and 31C, with C losing.  If so, have to finish 
> with A vs B.

My impression was that it was a hack - a way of getting a "summable" 
method that can be done using IRV voting machines and that's also at 
least slightly IRV-ish.

>>> The federal-state voting system certification process is a mess and
>>> the entire voting machine industry is a mess because they use
>>> proprietary standards and so voting system component are not
>>> interoperable, and the flawed design of voting machines makes it
>>> extremely difficult to check to see if the systems are producing
>>> accurate vote counts or not.
>>
>> To the extent of my knowledge, I agree. I think that having the 
>> machines be engineered around a summable method would help a lot - 
>> then the machines could be, to quote someone whose name escapes me at 
>> the moment, "expensive pencils". A Condorcet counting machine simply 
>> has to do the very simple job of iterating through the ranks; a Range 
>> counting machine  just has to turn optical scan configurations into 
>> numbers ("he filled in three circles of ten for candidate X" to "X: 
>> 3/10"). You're left with a small amount of information - the sum of 
>> the array or matrix - that can be made public.
> 
> Important goal here is letting voters express their desires, and having 
> this properly influence who gets elected.
>      Plurality is weak on the letting.
>      Approval is better, but gets proper complaints.
>      Condorcet and score do better, but duel as to which is better.
>      IRV allows about the same expression as Condorcet, but can deliver 
> embarrassing results.

For individual ballots, rated ballots probably confer the greatest 
freedom. One can simulate a ranked ballot (A: 9, B: 8, C: 7) with or 
without ties, and approval style (A: 10, B: 10, C: 0) or Plurality style 
as well.

That is, even though the voters may supply ranked or Approval type 
ballots (depending on the system in question), all of those could be 
stored as rated ballots, which means that in the event of having to 
change the voting method, the format doesn't change.

That's for individual ballots. But if the method is summable, one may go 
further and ask, what kind of summable data format would provide the 
most information? What kind would let one run as many different voting 
methods as possible without having to alter the format of what's being 
communicated from precincts and summed centrally? One candidate is the 
Condorcet matrix. Another is the weighted positional system matrix; or 
one may have both, to use methods like "first preference Copeland".

>> That may sound more vaporware-ish, but what I'm trying to say is that: 
>> if we could only have one change, let that be that rank order machines 
>> use a format that is summable and can be used by a variety of methods. 
>> Condorcet matrices fit (most Condorcet methods only need the matrix). 
>> Weighted positional methods (Borda, etc) can all use another kind of 
>> matrix ({x,y} contains the number equal to how many times candidate x 
>> was voted in yth place). And so on... what you have to decide is what 
>> format to use.
> 
> Cycles can occur with Condorcet, so the method must be able to do well 
> when there is no CW.

Yes. As long as the format is a Condorcet matrix, you may use any method 
that does well when there is no CW, as long as that method only required 
the Condorcet matrix. For instance, it would be quite easy to shift from 
Schulze to MAM/Ranked Pairs, since both work using the Condorcet matrix 
alone. In effect, one decouples the calculation (determining the 
winners) from the counting (determining what people actually voted), and 
one can thus alter one without necessarily having to alter the other.

(By the way, your From address was somewhat strange, purporting to be 
from my own ISP, so I used an older one - I hope it's the right one.)