[EM] No. Condorcet and Hare do not share the same problem with computational complexity and process transparency.

[Kristofer Munsterhjelm]

On 2024-03-18 02:03, Rob Lanphier wrote:
> Hi Kristofer,
> 
> I have a detailed reply below.  In short, I'm still pretty sure Michael 
> Ossipoff is worth listening to every so often (even though many of his 
> emails are thoughtless stream-of-consciousness that would get him banned 
> in most places, and I haven't ruled that out if it becomes clear he's 
> driving people away). 

That may be, but I feel he's rather too irascible to deal with, and that 
he gets his partisan preferences in the way of discussing methods. 
(Other readers, feel free to skip to "voting method stuff below".) 
Here's from the discussion that ultimately led to the plonkage:

On 2023-09-21, Mike argued in favor of IRV by (as I understood it) 
essentially saying that, given that IRV has compromising failure, any 
electorate that knew this and still went for IRV were tough enough not 
to compromise to begin with. The reasoning went that, as they know of 
IRV's compromising failure, they wouldn't choose a method that had 
compromising failure unless they were determined to avoid triggering 
that failure. Quoting:

> So I’m sure that I’ll propose & recommend good Condorcet versions
> (even if I don’t yet know which ones & how many) over IRV.  
> 
> …but I’ll nonetheless include IRV among the methods that I offer,
> because it’s better than a lot of people believe.   …though its merit &
> workability strongly depend on the electorate & the candidate-lineup.
> 
> I.e. Because it isn’t Condorcet-complying, it’s necessary that the
> electorate aren’t timid lesser-evil giveaway voters.
> 
> But an electorate that has just enacted IRV in a referendum didn’t
> do so because they want to rank Lesser-Evil over their favorite. They
> enacted it because they want to rank sincerely, to express & fully help
> their favorite.

Emphasis on the last sentence. Source 
http://lists.electorama.com/pipermail/election-methods-electorama.com/2023-September/004912.html. 
At the time I found this very strange, and it seemed to me that using 
such reasoning could lead to absurdity.

Then, on the 25th., he said that Coombs had too much of a burial 
incentive to be useful.

> Some academic authors have high praise for Coombs. One say that, with
> sincere ranking, & fewer than 5 candidates, Coombs always elects the CW.
> 
> But Coombs is obviously vulnerable to east burial strategy. In particular,
> trust & betrayal perpetrated by the voters of a “ lesser”-evil.
> 
> Sure, after that betrayal, they’d hopefully never have any support from
> their victims again.
> 
> But 1) Again we’re talking about resolution at least an election-cycle
> later; & 2) It could devolve to never-ending routine mutual burial.
> 
> Coombs doesn’t sound very promising to me.

Source: 
http://lists.electorama.com/pipermail/election-methods-electorama.com/2023-September/004941.html

So I thought I would respond by poking a bit of fun at it, and 
indirectly show how the IRV reasoning proved too much and could lead to 
absurdity:

>> But Coombs is obviously vulnerable to east burial strategy. In 
>> particular, trust & betrayal perpetrated by the voters of a “ lesser”-evil.
> 
> Clearly then, knowing this fact, the voters who propose and enact Coombs 
> must be tough voters who would never ever bury. Therefore Coombs' burial 
> incentive is no problem wherever it would be proposed.
> 
> I jest :-)

Apparently he got quite offended. He responded:

>> I jest :-)
> 
> …
> 
> …&, by so doing, you waste our time, & space at the postings-page, & send
> your substandard attempt at humor to everyone’s e-mail.
> 
> …
> 
> You’re aware that your bullshit is going to the e-mail of every
> list-subscriber, right?

and

> The only absurdity is in his sloppy attempt at an analogy, which has
> nothing in common with what it’s supposed to be an analogy for.   …& his
> equally sloppy & absurd conclusion from it (which he expressed as a serious
> “real point”, rather than as “jest”);

I enjoy these sorts of replies about as much as the next guy, which is 
to say not at all, so that was that. Now, he did say in his post that

> Perhaps Kristofer didn’t read my posts that said that RCV’s
> disadvantage is that it strongly depends on the electorate not being timid
> lesser-evil giveaway-voters.

which suggests that his point was not so much "electorates who propose 
IRV must necessarily have precommitted themselves to not do 
compromising" as "IRV will fail if the electorate hasn't". But if so, 
there are definitely better ways to to say "I think your joke is off the 
mark, you must have misunderstood".

I usually don't poke fun to press the absurdity of a point, so I think 
he had already got under my skin at that point. All the more reason to 
stay away.

Anyway,

=== voting method stuff below ===

> Credible voter models show that approval voting 
> and Condorcet consistency are practically compatible, even if they 
> aren't strictly compatible.  A system that "computers can count, even if 
> people can't" is not viable in our lifetimes, because people are more 
> complicated than computers.

The example I like to use here is Meek in New Zealand local elections. 
Meek's method uses a fixed point iteration to determine the keep values, 
and thus necessarily has to be counted by computer. I doubt you could go 
to an average New Zealand voter and get them to explain how Meek works. 
Yet they use it, so it's possible for the voters to trust a method with 
computerized counting.

And I suppose that's the point: trust. It's harder to trust a 
computerized system when it hasn't built up a reputation for good 
results, or when previous complicated systems have failed (if IRV is to 
be considered both a complicated system and one that failed).

Another example would be the Huntington-Hill apportionment method. It's 
not just complex but needlessly so (Webster would be better). I suspect 
the average voter would be hard pressed to explain how it works. Over 
here in Norway we also have a greedy algorithm that handles top-up 
leveling seats to improve national proportionality while also 
maintaining regional proportionality. Again, I doubt that an average 
voter could explain how it works; but they mostly trust it, so there's 
little problem. (Bizarre outcomes notwithstanding: personally I'd favor 
a change of algorithm, but that's another matter.)


I agree that Approval wins by a mile in the bang for the buck category. 
If your summability is restricted to one number per candidate, 
Approval/Range is the best you can get, but mostly because the other 
contenders make it no contest. But I can't shake the "manual DSV" and 
rb-j objections, that plain honest voters will be annoyed that they have 
to collapse their expressions into "yay? or boo?", and that the small 
risk of disastrous returns from misjudged strategy will eventually blow up.

Admittedly, I have no proof of this, since Approval hasn't been used 
much. I just know that's how I would think if my area switched to 
Approval. (We don't actually have single-winner elections, but you get 
my point :-)

>> I've got Mike plonked, so I don't see his posts,
> 
> 
> That's too bad.  Michael is frequently annoying, but he's frequently 
> correct.  This mailing list was started in large part because of a 
> mailing-list conversation I had with Michael in 1995, where he was being 
> obnoxious on another list.  I thought I'd be able to show that he was a 
> crank.  Turns out he taught me about center squeeze.  You should 
> consider unplonking him.

There are definitely things I disagree with him about, and that I would 
tell him were he, say, Forest. But I don't fancy getting my head bitten 
off again. Maybe I will, but I'm not sure yet.

>>     but I would like to add this:
>> 
>>     - If a lack of summability is not a problem, then BTR-IRV isn't that
>>     much more complex than IRV. And at the cost of slightly more complexity
>>     than that, Benham can preserve IRV's strategy resistance and do away
>>     with most of its exit incentive.
> 
> 
> Having volunteered as a poll worker for the first time in a city that 
> uses RCV for some elections, it changed my perspective on election 
> security.  I appreciated how much process there was, but also how much 
> of the process was shrugged off when it was a little inconvenient.
> 
> There weren't any RCV races in the March 5 election here, so I didn't 
> have to perform any tech support for RCV, but having voted in many RCV 
> races, I could see what a goat rodeo that can become for poll workers.  
> My hunch is that the more complicated the election, the easier it would 
> be to have steps of the process shrugged off as poll workers get 
> frazzled as the day wears on.
> 
> I think "summability" is really just shorthand for "vaguely makes sense 
> for someone who really really cares about the end result to keep track 
> of the election in real time".  Strict Condorcet methods are admittedly 
> difficult on this count.  Approval is drop-dead simple on this count.

That's more or less what I've come to think too. Summability primarily 
is about interpretability, and secondarily that people transporting the 
data have a chance to see if it's been tampered with. In both cases it's 
related to transparency.

Computer wise, it's possible to store full rankings for a worldwide 
election and a reasonable number of candidates on an SD card, even with 
100% turnout of 8.1 billion voters. So summability is not for storage 
purposes alone, unless you're doing a manual count (which again ties 
into transparency).

The interpretability point is weakened as you go from first order to 
second to third... and by the time you're doing real-time IRV sankey 
diagrams, all meaning is lost.

>> If computers do the counting, then relatively laborious steps aren't
>> any problem, as long as the public understands why they're there. 
> 
> I think that's an easy thing for those of us who are good with computers 
> to say.  Computers are taking over the world, but there's a limit to how 
> much people trust computers and the people who write the software for 
> computers.  Many people "trust" computers only as far as they can throw 
> a datacenter.  Granted, it's possible to wire up many computers in a 
> small box that most healthy adults can throw and call that a 
> "datacenter", but I'm talking about the brick-and-mortar datacenters 
> often placed near power generation plants.  Most people have given up 
> the fight, and welcome our robot overlords, but our robot overlords 
> don't really care if we understand elections, and may prefer to do away 
> with elections and take control themselves.  :-)

I'm definitely not going to propose that large language models call 
elections :-)

> In seriousness, I'm guessing this mailing list skews heavily "math 
> literate" in addition to skewing heavily "computer literate", and I 
> think that those of us that are literate in those way have a hard time 
> relating to people that aren't as literate in those areas:
> https://www.washingtonpost.com/outlook/math-hard-easy-teaching-instruction/2021/06/25/4fbec7ac-d46b-11eb-ae54-515e2f63d37d_story.html <https://www.washingtonpost.com/outlook/math-hard-easy-teaching-instruction/2021/06/25/4fbec7ac-d46b-11eb-ae54-515e2f63d37d_story.html>

And that's a good point: the curse of knowledge is very real.

>>    BTR-IRV's
>>    safeguarding step follows directly from your concept that "if more
>>    people prefer A to B than vice versa, then B must not be elected".
> 
>>    - If, on the other hand, lack of summability *is* a problem, then that
>>    disqualifies IRV outright and we're done.
> 
> 
> I'm supportive of BTR-IRV, but I'll concede that summability and 
> reporting results in an easy-to-understand form (in real time) is a big 
> problem.  I think it's important for voters (on election night) to be 
> able to see a television reporter say "Results from the precincts on the 
> southwest side of town were just reported, and CandB took the lead over 
> CandA.  Let's turn it over to our analysts at the elections desk to 
> explain what happened!"  The pre-election polling and exit polling 
> should provide a reasonably understandable explanation.  I fear we're 
> due for a lot of election fraud if most people don't understand what 
> happened (and honestly, having lived in San Francisco since 2011 and 
> seen how some close elections have turned out, it wouldn't surprise me 
> at all if there's some consequential electoral fraud here).

I'm no fan of IRV either. I'm just saying "if IRV, then Condorcet-IRV". 
That is, the return on including some Condorcet provision if you're 
going to do IRV anyway is high enough that you really ought to do it.

As far as reporting goes: does anyone here know how Australia does it?

I suppose their above-the-line voting makes IRV much more like largest 
remainders party list, but I've heard that optional voting is becoming 
more common, which could lead more voters to manually rank the candidates.

I also have the impression that polls include pairwise data 
("two-party-preferred") showing the relative support between the two 
party blocs (Labour and LibNat). But I don't want to mess up the 
details, so I'll leave them to someone who actually lives in Australia.

French presidential polls seem to include hypothetical expected pairwise 
results between the candidates who have some chance of getting into the 
final. See 
https://www.politico.eu/article/5-charts-to-help-you-read-the-french-presidential-election/, 
figure near the bottom. So if we look more broadly, pairwise reporting 
isn't completely unheard of.

> I hear you, and I read what Forest wrote.  Ultimately, I think it's 
> important for most voters to vaguely know what the election is going to 
> look like in order to be comfortable using the system.  I don't think 
> most folks here in the SF Bay Area really understand RCV.  The topic 
> frequently comes up on the nightly news, for example here:
> https://www.ktvu.com/news/lawsuit-filed-to-overturn-oakland-mayoral-election <https://www.ktvu.com/news/lawsuit-filed-to-overturn-oakland-mayoral-election>
> My fear is that RCV makes fraud easier, because few people truly 
> understand what's going on under the hood, and the founders of FairVote 
> don't help educate; they obfuscate.  I'm hopeful that FairVote will get 
> over their nasty case of "founder's syndrome" soon, so that they will 
> become better partners in electoral reform efforts.

IRV RCV is entirely nonsummable. (Summable) Condorcet should do better, 
particularly in the absence of cycles. One could possibly do reporting 
by saying something like "candidate X is still the champ, but his match 
against candidate Y is evening out - what does that mean, is this region 
a stronghold of Y's? Did the voters here prefer Y to X because of 
economic reasons?", etc.

When things get cyclical it gets a lot tougher. But simple rules could 
possibly still work, e.g. minmax as "your strength is the strength of 
the most unfavorable matchup". Reporting could talk about how X's 
apparent comfortable margin is taking a beating on some issue that 
candidates Y and Z are strong at, and that X's victory is looking slim 
because Y is already doing a good job in the X vs Y contest. Who the 
current champ is and how his winning chances are either being shored up 
or eroded by more votes coming in.

Copeland is probably quite easy to understand although indecisive and 
not cloneproof. Brackets could be simple, but I don't know of any 
Condorcet method that uses them -- and the seed order would have to be 
set in advance. Otherwise, as more votes come in, it could alter the 
seed order and make the comparisons seem like an unpredictable mess.

>> [1] Both honest voters in the rank-consistent sense and in the von
>> Neumann-Morgenstern sense.
> 
> 
> Could you explain what you mean by this?

What I mean is that both honest voters who have a particular rating in 
mind, but not a ranking; and honest voters who have strengths of 
preference in mind, have multiple honest ballots to choose between. So 
the choice of which one to use becomes a matter of what others are 
doing, even for people who would rather not do strategy.

This is most obvious for ranked voters: if your opinion is A>B>C, you 
don't know if you should approve only A or both A and B.

von Neumann-Morgenstern utilities is a way to quantify strength of 
preference by using lotteries and expected utility.

Suppose that your preference is A>B>C, and that you think that getting B 
for sure is as good as a gamble with a 40% chance of A, and a 60% of C. 
Then your von Neumann-Morgenstern utility for B is 40% of the one for A 
plus 60% of the one for C. E.g. if your rating of C is zero and A is 10, 
then B is 4.

By considering what gamble you would find about as good as getting a 
candidate for sure, you can (theoretically) determine your strength of 
preference for all other candidates once you have two of them. So that 
allows a more meaningful theory about what strength of preference really 
is, and to say that a ballot is honest if it's consistent with these 
preference strengths.

But there's still a problem: you're left with two free variables - the 
ratings of your favorite and least favorite. So there are still multiple 
honest Range ballots. And if we suppose that Approval works by approving 
every candidate above the halfway point on the rating scale, then there 
are still multiple honest Approval ballots, too.

We could get around this by fixing the voter's favorite candidate to a 
rating of 100% and the voter's least favorite to 0%. Now there is only 
one honest rated-like ballot. But methods that automatically normalize 
like this fail IIA, and both in Range and (above mean utility) Approval, 
there can exist an incentive to not cast that honest ballot.

(In practice, people don't like risk and so prefer a sure deal over a 
gamble, but there are ways to compensate for this too.  The point is 
that it provides a formalization of the idea of "strength of preference".)

-km