[EM] Dyadic Approval definition (continued)

[Forest Simmons]

Hi, EM list folks. I decided to make this continuation cumulative so that
we can have the whole thing in one message (eventually).

As in Rocky and Bullwinkle's "Here's what happened in our last episode",
there may be a few slight improvements in this version of the first
installment. 

First Installment:

Dyadic Approval is a generalization of both Approval and Condorcet.

It generalizes Condorcet by allowing expression of degrees of preference
(in a consistent way that doesn't encourage preference reversals or
collapses) as well as by retaining the pairwise aspect of Condorcet. 

It generalizes Approval by allowing the voter to make refinements within
various previously defined categories of Approval by specifying which
candidates within a category would be approved if they were the only ones
left. 

I find it is easiest to start with a sample voted ballot, and then
interpret it, and explain how it is scored and put into matrix form for
addition into the election totals.

Once the total matrix is in, the winner can be determined from the matrix
by SSD or any other good Condorcet method. However I am going to recommend
another method that takes advantage of the hierarchical structure of the
ballots for use when there are (generalized) Condorcet cycles (cycles of
pairwise preferences among candidates).

Here's a sample voted ballot with eight candidates:

        A > B >> C > D >>> E > F >> G > H

This one is ideal ... a precisely balanced binary tree with one leaf at
the end of each twig. I'll show some more typical and extreme ones
presently, to demonstrate the range of possibilities.

We interpret this ballot as follows. The voter supports (on the coarse
level of ordinary Approval) candidates A through D, and rejects E through
H.

However, if candidates A through D were to be eliminated (by
assassination, Approval runoff, heart attacks, etc.) then among the
remaining candidates E and F would be supported, while G and H would be
disapproved. Similarly, if all of the candidates except C and D were
eliminated, then the voter would approve only C.  I think you can see the
pattern.

For scoring, we note that the coarsest distinction is made by a string of
three inequality signs: >>>, so the score of any pairwise comparison of a
pair that straddles that string receives a weight of 1. Then any pair that
straddles >> (but not >>> ) receives a weight of one half.  Similarly, any
pairwise comparison between candidates on opposite sides of a single >
symbol (but no string of them) receives a weight of one fourth. 

Another voted ballot (for seven candidates this time) could look like the
following: 

      A >> B > C >>> D >>>> E, F >>> G

In this ballot the voter chose to not distinguish between E and F.  All
comparisons between E and F add zeros to the corresponding positions
in the scoring matrix. Furthermore, the underlying binary tree is not as
well balanced; some possible categories are missing.

The coarsest comparisons are the ones straddling the string of four
inequality signs ( >>>> ) so they are the ones that receive a weight of
one on this ballot.  By the time we get down to the comparison between B
and C, it only gets a weight of one eighth.

In particular, the comparison between the pair A and G  gets the same
weight as the comparison between D and E, so there is no temptation to
insincerely move D further up the list or E further down the list.

Two extreme possibilities are:

     A > B >> C >>> D >>>> E >>>>> F >>>>>> G

                       and

     A >>>>>> B >>>>> C >>>> D >>> E >> F > G

In the first of these two extreme cases, all comparisons relative to G
have a weight of one. Only the comparison between A and B has a weight of
1/32 . If this were a zero info election, it would mean the utilities were
skewed towards the upper end of the scale. 

In the second case, all comparisons relative to A receive a weight of one,
while the comparison between F and G has a weight of 1/32 .  If this were
a zero info election, it would mean the utilities were skewed toward the
low end of the scale. 

Promises for future installments:

(1) the binary string representation of the ballots.

(2) natural hierarchy resolution of cycles.

(3) answers to frequently asked questions.

(4) how to use these same style ballots for Approval Runoff.

--------------------------------------------------------------------

Second Installment:

There are two things I would really like to emphasize to this point.

First, the Dyadic Approval ballot is potentially richer in information
than either a plain preference ballot or a plain Approval ballot.

To convert to a plain preference ballot, just ignore the fact that there
are several kinds of inequality strings, and treat them all the same (i.e.
as in Condorcet). 

To recover the plain Approval information, just ignore all the approval
refinements (i.e. all the approval levels except the coarsest).

Second, if the ballots are only used to compare the candidates pairwise
(with consistent weights as outlined above), then there is no reason for
any voter to reverse any preference or to collapse any preference.

By collapse a preference, I mean making no distinction when there really
is a preference.  Regular Approval falls short in this regard because the
voter has no choice but to collapse some preferences. Cardinal Ratings
(CR) falls short because there are strong strategic incentives to vote
only at the extremes.

Putting these two observations together, we see that Dyadic Approval not
only has more information potential than either plain Approval or plain
Condorcet, it actually realizes that potential (when scored
appropriately). The voted ballots are indeed richer in information than
either plain Approval or plain Condorcet voted ballots. 

To understand the significance of this, consider the case of CR. The CR
ballot has potential for more information than either Approval or
Condorcet, but it cannot realize that potential (except on the perfect
honor system) since there are strong strategic incentives to inflate and
deflate the ratings all of the way to the respective extremes. 

As near as I know, Dyadic Approval and its obvious variations are the only
high resolution methods that don't suffer a strategic collapse of
information. (Compare Borda, a rather low resolution method that suffers
from strategic order reversals, worse than the mere collapses of CR and
plain Approval.)


On a practical note, a real Dyadic Approval election with several
candidates (and unsophisticated voters) would have to be via interactive
voting machines that that repeatedly ask the voters, "Which of THESE
candidates would you support if THEY were the only ones in the race?"  At
any point the voter could decide to start all over again, so the line at
the polls might get a little long. 

For sophisticated voters in a committee situation, a voted ballot would be
written vertically (to make room for longer names) instead of
horizontally, and the marks separating the categories would be, lines,
double lines, triple lines, etc.

[Sorry this has to be the end of the short second installment.]