[EM] Sequential Pairwise Elimination versus Sequential Loser Elimination

classic Classic list List threaded Threaded
1 message Options
Reply | Threaded
Open this post in threaded view

[EM] Sequential Pairwise Elimination versus Sequential Loser Elimination

Forest Simmons

From: Forest Simmons <[hidden email]>

This is a continuation of the above named topic/subject discussion starting with an example showing Chicken Defense compliance ...

Here's the scenario with sincere preferences in brackets:

49 C
26 A>B
25 B [sincere B>A]

Implicit Approval plus least minimum support:

A 26 + 26
B 51 + 25 (B against A)
C 49 + 49

So the seed/agenda order is A(52), B(76), C(98)

The pairwise defeat cyclic order is A beats B beats C beats A.

So the sequential pairwise single elimination tournament summary is

B is eliminated by A, which in turn is beaten by C, the Sequential Pairwise Elimination winner.

As we now show, the B faction would have done better by voting sincerely:

In that case the only change in seed score would be A's, from 52 to 77, due to an increase in A's approval, so the line up would be

B, A, C 

But this time in the second contest A beats C, 51 to 49, so the sincere CW wins which is indeed better for B.

What if, the sincere preferences [in brackets] were ...

49 C [C>B]
26 A>B
25 B [B>C]

In this case B is the sincere CW: B beats A, 74 to 26, and B beats C, 51 to 49.

Evidently C has attacked B by truncation (partial burial ... full burial would play out the same way), while the B faction has employed the standard CW prudent defense of truncation below the sincere CW.

We have just seen that the Chicken Defense Criterion requires C or A to win in this configuration, but the Plurality criterion requires A to lose. It seems that the same ballot configuration requires C to win to satisfy Chicken Defense and Plurality, and for C to lose in order to defend against (i.e. not reward) a truncation or burial attack by C.

Fortunately, the natural way out of this dilemma is for the A faction to pursue its best interest by raising its compromise B to equal first. Then B becomes the ballot Condorcet winner, the implicit winner, and the candidate whose min pairwise support is greatest, all at once!

In sum, in the face of rational voters pursuing their own interests, voters can defend against burial of the sincere CW without any ballot rank reversals. 

Contrast this with IRV where voters can save their compromise fall back sincere CW candidate from the center squeeze effect (so prominently ubiquitous in IRV Yee diagrams) only by blatant insincere order reversals that painfully violate the conscience of the honest voter! 

Or for that matter, how would the sincere CW candidate B fare with sincere ballots in the 2nd scenario above? She would be eliminated in the first round ... with no attack from the C faction!
. under zero information conditions! No way to defend!

Under perfect information the A faction could save its compromise B by insincerdly burying favorite A under B... ouch!!!

Isn't election reform supposed to save us from lesser-evilism? (as much as possible)

I hope you find this explanation to be helpful. It is intended to help leaders understand the stakes, and give them enough understanding to see why zero strategy rankings are optimal for SPE (as with IRV) and that SPE defense against complete information manipulative attacks are thwarted with greater finesse (with less violence to the conscience) compared to Voting under IRV.

True, no method is perfect, but some methods are much better than others on the qualities that matter most!

Is there any good property of IRV that is more important to you than the properties we have discussed so far? How about Later No Help/Harm?  Are you willing to trade monotonicity and the Condorcet Criterion for them?

Let's get all these concerns out in the open!

I have suggested that the compliances of SPE are the most important ones ... sincere strategy optimality under zero info conditions including no accidental eliminatiln of the Condorcet Candidate,  ease of defending the CC when under strategic attack, efficient precinct summability and transparency of the count, Clone Independence, Monotonicity, Plurality, and Immunity to Chicken Attack.

The most important missing criterion is the FBC, which seems to be incompatible wth the Condorcet Criterion....but our SPE has a base method for the agenda/seeding that satsfies the FBC .. giving max FBC compliance for a Condorcet method. In fact, if you skip the elimination steps and jump straight to the last (i.e. most promising) candidate in the lineup/agenda you get a perfectly good FBC compliant method ... the Approval Stable Winner ... the candidate whose sum of Implicit Approval and
Minimum Support against any other candidate is maximal.

I'll take a break here ...
---------- Forwarded message ----------
From: Forest Simmons <fsimmons@for pcc.edu>
Date: Saturday, February 27, 2021
Subject: Condorcet method - Wikipedia

The Wikipedia Article [link below] on Condorcet Voting describes Sequential Pairwise Elimination as the most studied version of Condorcet voting in the literature because it is the one most widely used in deliberative assemblies as recommended in Robert's Rules of Order.

It is important to distinguish Sequential Pairwise Elimination from (mere) Sequential Loser-Elimination methods like Raynaud, IPE, IRV, etc. that may or may not use pairwise information in determining what kind of loser to eliminate at each step. The crucial difference between the two kinds of elimination (SPE and SLE) in the case where they both require a pairwise loss in the elimination step is that the former follows a predetermined schedule or agenda (the seeding order in a single elimination tournament) of whom to compare next with the winner of the previous round (the champion so far) while the latter determines the next pair of candidates to be compared pairwise as though starting from scratch, with all remaining (uneliminated) candidates on an equal footing.

The difference might seem unimportant but it is essential for monotonicity criterion compliance, which elimination methods invariably fail when not working off a monotonic agenda. This compliance is a clear advantage over IRV that we cannot afford to throw away.

For example, some have suggested that Borda be used as the base method for SLE ...at each step eliminate the remaininng candidate with the lowest Borda score (Baldwin's method)... if we wanted to reinforce Condorcet compliance we could change that to eliminating the pairwise loser between the two lowest Borda score candidates ... but are we talking Borda scores as determined at the beginning? ...or newly calculated Borda scores reflecting only the rankings relative to the remaining candidates?

Of these options only the SPE version (the one with the Borda agenda adhered to throughout) is monotone. I use it only as a example to clarify the difference between SPE and SLE.

Why not propose it? It beats IRV on four important counts ... monotonicity, Condorcet compliance, efficient precinct summability, and transparency. It is equally easy to vote since it makes use of the same ranked preference ballots.  Why not propose it?

Because IRV is better in the single most important way ... IRV is clone independent.  Clone Idependence (with sincere zero information voting) is the only justification for proposing any method requiring a ranked preference ballot.... IRV and various other methods pass on this score, but all Borda based methods fail ... it's a deal breaker.

So what should we use as an agenda setter (seeding order) if not Borda?  Why not Approval? Approval would provide a monotone, clone free, agenda order, so why not propose SPE Approval?

One tiny objection  ... how to decide approval cutoffs ... an additional burden on the voter (frown emoji).

Suggested solutions include....

... making approval marks optional with default approval just above truncation or just below equal top, or the average of the two... not a bad idea, though slightly complicating the count and worrying the voters about the option ... (If I don't make best use of the option, am I failing my civic duty?)

....Candidate Proxy where voters mark the approvals and disapprovals that they feel sure about and delegate the rest to their favorite or other candidates to decide.... great idea, but not worth the miniscule increase in complexity in the minds of some.

.... DSV Approval. A good DSV Approval cutoff would be the lowest candidate on the ballot that pairwise beats all candidates ranked above it. But this complicates the count ... one pass to determine the pairwise matrix, and another to determine the approvals for the agenda/seed order.

...Imlicit Approval would not work in Australia where complete ballot rankings are mandatory (no truncations or equal rankings allowed).

How about seeding the candidates from least minimum pairwise support to greaest minimum pairwise support? This works best when the pairwise support of X against Y is defiined as the number of ballots on which X is ranked greater than or equal to Y.

Or how about the "stable approval" order,  a synergistic sum of these last two suggestions that optimizes zero information strategy ... the best strategy when no reliable horse race information is available is to vote sincere ranked preferences as you would under IRV under the same zero info conditions.

With this seeding of the agenda, our SPE method also satisfies Plurality and Immunity to "Chicken" threats, two important criteria that IRV supporters could embarrass us with if we failed them.

When I get another free moment I will give a couple of examples ... for now I conclude with the following observation ...

Remember, the support for X against Y is the number of ballots on which X is voted above or equal to Y ... which is identical to X's approval score, no matter the choice of Y,  whenever all voters vote only at the extremes ... no intermediate ranks. So in that case there is no difference in seeding by Implicit Approval (least to most) or seeding by minimum pairwise support (least to most). Our proposed SPE method reduces to approval whenever no intermediate ranks are occuppied on the ballot: This is a longstanding tradition for acceptable EM list methods ...

In particular the following example that shows Plurality failure for the MMPO (MaxMin Paiirwise Opposition) method elects the approval winner:

44 A=C
13 C
43 B=C

awards the win to the Approval winner A:

The seeding order is C, B, A while the pairwise beat order is A beats B beats C, so C loses to B who in turn loses to A ... the undisputed champ!

 To be continued ....

---------- Forwarded message ----------
From: Susan Simmons <[hidden email]>
Date: Saturday, February 27, 2021
Subject: Condorcet method - Wikipedia
To: Forest Simmons <[hidden email]>


Sent from my MetroPCS 4G LTE Android Device

Election-Methods mailing list - see https://electorama.com/em for list info