Discussion:
MinMax variant
Forest Simmons
2003-03-07 18:42:55 UTC
Permalink
Has anybody ever proposed minimizing the maximum opposition rather than
minimizing the maximum defeat?

I know that theoretically this could elect the Condorcet Loser, but it
seems very unlikely that it would do so.

It seems to me that if equality were allowed in the rankings, then this
method would satisfy the FBC, since ranking Favorite equal with Compromise
wouldn't increase opposition against Compromise.

Venzke Kevin has recently suggested a version of this method adapted to
approval ballots.

I wonder how well it would work with CR ballots having resolution greater
than two.

You would start with the pairwise matrix whose (i,j) entry is the number
of ballots expressing a preference for candidate i over candidate j.

Then you would find the maximum entry in each column.

Then the minimum of these numbers corresponds to the candidate whose
maximum opposition is minimum.

This method differs from the usual MinMax in that it considers all
opposition, not just opposition resulting in defeat.

Forest
Steve Eppley
2003-03-07 23:07:48 UTC
Permalink
Post by Forest Simmons
Has anybody ever proposed minimizing the maximum opposition
rather than minimizing the maximum defeat?
It's a reasonably good method, although it doesn't satisfy some
criteria I consider important such as Minimal Defense (which is
similar to Mike Ossipoff's Strong Defensive Strategy Criterion) and
Clone Independence. (That's why I think the best method is a
variation of Ranked Pairs which I call Maximize Affirmed Majorities,
or MAM. See the web pages at www.alumni.caltech.edu/~seppley for
more info and rigorous proofs. The website is still under
construction, not yet friendly for people who aren't social
scientists, and most of the web pages requires a web browser that
supports HTML 4.0 and Microsoft's "symbol" font to be viewed
properly.)

Minimax(pairwise opposition) even satisfies a criterion promoted by
some advocates of Instant Runoff, which I call "Uncompromising":

Let w denote the winning alternative given some set
of ballots. If one or more ballots that had only w
higher than bottom are changed so some other
"compromise" alternative x is raised to second
place (still below w but raised over all the other
alternatives) then w must still win.

The proof that Minimax(pairwise opposition) satisfies Uncompromising
is simple: Raising x increases the pairwise opposition for all
candidates except w and x, and does not decrease the pairwise
opposition for any candidate, so w must still have the smallest
maximum pairwise opposition.

That criterion can be strengthened somewhat and still be satisfied:
Changing pairwise indifferences to strict preferences in ballots that
ranked w top cannot increase w's pairwise opposition or decrease any
other alternatives' pairwise opposition.
Post by Forest Simmons
I know that theoretically this could elect the Condorcet Loser, but it
seems very unlikely that it would do so.
-snip-

Minimax(pairwise defeat) can also elect a Condorcet Loser. Suppose
there are 4 alternatives, 3 of them in the top cycle but involved in
a "vicious" cycle. If the pairwise defeats of the 4th are slim
majorities, the 4th wins.

As for the likelihood, this may depend on how you model voters'
preferences. In a spatial model with sincere voting, I think you're
right. But what if supporters of the Condorcet Loser vote
strategically to create a vicious cycle?

Minimax(pairwise opposition) can also defeat a "weak" Condorcet
Winner, one that wins all its pairings but does not have more than
half the votes in each of its pairings. The CW may defeat candidate
x pairwise by a plurality such as 48% (less than half the votes) and
that might be x's largest opposition, whereas the CW may have a
larger opposition, such as 49%, in some pairing.

But it satisfies a weaker Condorcet-consistency criterion:

If there is a "strong" Condorcet Winner (an alternative
that wins each of its pairings by more than half of the
votes) then it must be elected.

Thus anyone who claims no Condorcet-consistent method satisfies
Uncompromising is incorrect.

-- Steve Eppley
Forest Simmons
2003-03-15 00:05:25 UTC
Permalink
How about this modification of this variant?

Start with the pairwise matrix P in which the number in the row
corresponding to candidate X and the column corresponding to candidate Y
is the number of ballots ranking (or rating or grading) X strictly above
Y.

Order the candidates according to the maximum entry in their columns, i.e.
their maximum pairwise oppositions. [Of the two possible orders, choose
the one that puts the max max opposition candidate at the bottom of the
list and the min max opposition candidate at the top.]

Bubble sort this order to get its local Kemenization.

This is done by starting at the top and working your way down through the
order, letting each candidate percolate up as far as possible by
transpositions of adjacent candidates whenever dictated by the pairwise
win matrix. [This pairwise win matrix is obtained from the pairwise matrix
P by subtracting its transpose from it, replacing all resulting positive
numbers by ones, and zeroing out the rest of the matrix.]

I like this method because it comes close to satisfying the FBC.

Why would you rank Compromise over Favorite? Only if you thought that
Favorite would get in the way of Compromise during the percolation
process, i.e. only if ...

(1) Favorite can beat Compromise head-to-head, and (2) Favorite's max
opposition is smaller than Compromise's max opposition, AND (3) Compromise
can beat all the candidates who at worst give up fewer points than
Favorite, including at least one that Favorite cannot beat.

You would have to be pretty fickle to abandon Favorite on account
of some poll pretending to be accurate enough to predict all of that.

Forest
Post by Steve Eppley
Post by Forest Simmons
Has anybody ever proposed minimizing the maximum opposition
rather than minimizing the maximum defeat?
It's a reasonably good method, although it doesn't satisfy some
criteria I consider important such as Minimal Defense (which is
similar to Mike Ossipoff's Strong Defensive Strategy Criterion) and
Clone Independence. (That's why I think the best method is a
variation of Ranked Pairs which I call Maximize Affirmed Majorities,
or MAM. See the web pages at www.alumni.caltech.edu/~seppley for
more info and rigorous proofs. The website is still under
construction, not yet friendly for people who aren't social
scientists, and most of the web pages requires a web browser that
supports HTML 4.0 and Microsoft's "symbol" font to be viewed
properly.)
Minimax(pairwise opposition) even satisfies a criterion promoted by
Let w denote the winning alternative given some set
of ballots. If one or more ballots that had only w
higher than bottom are changed so some other
"compromise" alternative x is raised to second
place (still below w but raised over all the other
alternatives) then w must still win.
The proof that Minimax(pairwise opposition) satisfies Uncompromising
is simple: Raising x increases the pairwise opposition for all
candidates except w and x, and does not decrease the pairwise
opposition for any candidate, so w must still have the smallest
maximum pairwise opposition.
Changing pairwise indifferences to strict preferences in ballots that
ranked w top cannot increase w's pairwise opposition or decrease any
other alternatives' pairwise opposition.
Post by Forest Simmons
I know that theoretically this could elect the Condorcet Loser, but it
seems very unlikely that it would do so.
-snip-
Minimax(pairwise defeat) can also elect a Condorcet Loser. Suppose
there are 4 alternatives, 3 of them in the top cycle but involved in
a "vicious" cycle. If the pairwise defeats of the 4th are slim
majorities, the 4th wins.
As for the likelihood, this may depend on how you model voters'
preferences. In a spatial model with sincere voting, I think you're
right. But what if supporters of the Condorcet Loser vote
strategically to create a vicious cycle?
Minimax(pairwise opposition) can also defeat a "weak" Condorcet
Winner, one that wins all its pairings but does not have more than
half the votes in each of its pairings. The CW may defeat candidate
x pairwise by a plurality such as 48% (less than half the votes) and
that might be x's largest opposition, whereas the CW may have a
larger opposition, such as 49%, in some pairing.
If there is a "strong" Condorcet Winner (an alternative
that wins each of its pairings by more than half of the
votes) then it must be elected.
Thus anyone who claims no Condorcet-consistent method satisfies
Uncompromising is incorrect.
-- Steve Eppley
_______________________________________________
Election-methods mailing list
http://lists.electorama.com/listinfo.cgi/election-methods-electorama.com
Markus Schulze
2003-03-08 01:31:10 UTC
Permalink
Dear Steve,
Post by Steve Eppley
Minimax(pairwise opposition) even satisfies a criterion promoted by
Let w denote the winning alternative given some set
of ballots. If one or more ballots that had only w
higher than bottom are changed so some other
"compromise" alternative x is raised to second
place (still below w but raised over all the other
alternatives) then w must still win.
The proof that Minimax(pairwise opposition) satisfies Uncompromising
is simple: Raising x increases the pairwise opposition for all
candidates except w and x, and does not decrease the pairwise
opposition for any candidate, so w must still have the smallest
maximum pairwise opposition.
Changing pairwise indifferences to strict preferences in ballots that
ranked w top cannot increase w's pairwise opposition or decrease any
other alternatives' pairwise opposition.
That sounds like Woodall's later-no-harm + later-no-help.
(Douglas R. Woodall, "Monotonicity of single-seat preferential election
rules," Discrete Applied Mathematics, vol. 77, pp. 81-98, 1997.)

In another paper, Woodall proves that no election method can
simultaneously meet later-no-harm, later-no-help, monotonicity,
and mutual majority. Therefore, the fact that Minimax(pairwise
opposition) violates mutual majority in such a drastic manner
can be considered a consequence of the fact that it meets
later-no-harm, later-no-help, and monotonicity.

Example:

20 ABCD
20 BCAD
20 CABD
13 DABC
13 DBCA
13 DCAB
Post by Steve Eppley
That's why I think the best method is a variation of Ranked Pairs
which I call Maximize Affirmed Majorities, or MAM.
In so far as you have always considered Mike Ossipoff to be
authoritative, I would like to know what you think about the
fact that he doesn't promote Ranked Pairs anymore at his
web pages.

Markus Schulze
Steve Eppley
2003-03-10 21:11:12 UTC
Permalink
Post by Markus Schulze
Post by Steve Eppley
That's why I think the best method is a variation of Ranked Pairs
which I call Maximize Affirmed Majorities, or MAM.
In so far as you have always considered Mike Ossipoff to be
authoritative, I would like to know what you think about the
fact that he doesn't promote Ranked Pairs anymore at his
web pages.
Mike Ossipoff has recently written to me that he considers
MAM to be the best voting method for committees and public
elections, which means his web pages (maintained by Russ
Paielli) are merely out of date.

I agree with Mike that MAM is best, but I don't agree with
him on everything. (For example, I have a much lower
opinion of Approval than he has.) So I don't understand
what Markus means when he claims I've always considered
Mike to be "authoritative." (Perhaps Markus should define
the threshold between authoritative and non-authoritative,
since he uses the term in an absolutist way.) I don't
consider Mike to be more authoritative than Markus or
myself.

See www.alumni.caltech.edu/~seppley for the definition of
MAM and rigorous proofs that it satisfies numerous
criteria.

-- Steve Eppley
Markus Schulze
2003-03-08 09:58:14 UTC
Permalink
Post by Markus Schulze
Post by Steve Eppley
Minimax(pairwise opposition) even satisfies a criterion promoted by
Let w denote the winning alternative given some set
of ballots. If one or more ballots that had only w
higher than bottom are changed so some other
"compromise" alternative x is raised to second
place (still below w but raised over all the other
alternatives) then w must still win.
The proof that Minimax(pairwise opposition) satisfies Uncompromising
is simple: Raising x increases the pairwise opposition for all
candidates except w and x, and does not decrease the pairwise
opposition for any candidate, so w must still have the smallest
maximum pairwise opposition.
Changing pairwise indifferences to strict preferences in ballots that
ranked w top cannot increase w's pairwise opposition or decrease any
other alternatives' pairwise opposition.
That sounds like Woodall's later-no-harm + later-no-help.
Sorry. Eppley's Uncompromising seems to be only Woodall's later-no-harm:
http://groups.yahoo.com/group/single-transferable-vote/files/Woodall_Monotonicity_DAM_1997.pdf

Markus Schulze
Markus Schulze
2003-03-09 02:02:53 UTC
Permalink
Dear Mike Ossipoff,

in the past, you have always considered Steve Eppley to be
authoritative. What do you think about the fact that he
promotes Ranked Pairs at his web pages? What do you think
about his arguments for Ranked Pairs against the beat path
method? What do you think about the fact that although he
uses a similar strategy concept he promotes a different
method than you do?

Markus Schulze
Markus Schulze
2003-03-07 21:50:46 UTC
Permalink
Dear Forest,
Post by Forest Simmons
Has anybody ever proposed minimizing the maximum opposition rather than
minimizing the maximum defeat? I know that theoretically this could elect
the Condorcet Loser, but it seems very unlikely that it would do so.
Of course, the fact that the MinMax method can elect the Condorcet Loser
has nothing to do with the way the strength of a pairwise comparison is
measured.

Example:

20 ABCD
20 BCAD
20 CABD
13 DABC
13 DBCA
13 DCAB

Markus Schulze
Forest Simmons
2003-03-15 00:16:02 UTC
Permalink
I had a simpler example in mind:

Three teams in a round robin tournament compile the following scores:

A beats B four to three, B beats C two to one, and A beats C two to one.

A is the Condorcet winner while C is not only the Condorcet loser, but
also the method winner under this variant of MinMax, since this team never
gave up more than one point per game, while each of the other teams had at
least two points scored against it in one game or another.

Forest
Post by Markus Schulze
Dear Forest,
Post by Forest Simmons
Has anybody ever proposed minimizing the maximum opposition rather than
minimizing the maximum defeat? I know that theoretically this could elect
the Condorcet Loser, but it seems very unlikely that it would do so.
Of course, the fact that the MinMax method can elect the Condorcet Loser
has nothing to do with the way the strength of a pairwise comparison is
measured.
20 ABCD
20 BCAD
20 CABD
13 DABC
13 DBCA
13 DCAB
Markus Schulze
_______________________________________________
Election-methods mailing list
http://lists.electorama.com/listinfo.cgi/election-methods-electorama.com
Forest Simmons
2003-03-15 01:44:52 UTC
Permalink
corrections:

"one" changed to "TWO", "two" changed to "THREE" in the last sentence
Post by Forest Simmons
A beats B four to three, B beats C two to one, and A beats C two to one.
A is the Condorcet winner while C is not only the Condorcet loser, but
also the method winner under this variant of MinMax, since this team never
gave up more than TWO points per game, while each of the other teams had
at
least THREE points scored against it in one game or another.
Continue reading on narkive:
Loading...