<div dir="ltr">lol</div><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Thu, Dec 2, 2021 at 12:06 PM Colin Champion <colin.champion@routemaster.app> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">To judge from the literature I would suppose that voting theory was part <br>
of first-order logic or of graph theory, but it seems clear that the <br>
right answer is Bayesian decision theory.<br>
<br>
This is easiest to see under a spatial model, but I think it’s perfectly <br>
general. We have vague prior information about the attributes of voters <br>
and candidates (eg. their positions in space) and about voter behaviour <br>
(how they will cast their ballots in the light of these attributes). We <br>
can condition this prior information on the observations contained in a <br>
set of ballots and thereby compute the posterior probability of any <br>
desired function of the voters' and candidates' attributes. Our aim is <br>
to identify the most representative candidate, where the degree of <br>
representativeness can be expressed through a loss function, and where, <br>
therefore, we will seek to identify the candidate whose posterior <br>
expected loss is the least. The prior knowledge of voters' positions can <br>
be thought of as a distribution of distributions, eg. the voters come <br>
from a mixture of three identical and equally weighted circular <br>
Gaussians whose means come from a further Gaussian.<br>
<br>
This is a constructive approach which in principle might be used to <br>
determine the winner of an election. We'd just need to integrate out all <br>
the unknowns to find the expected losses of the candidates. This was in <br>
essence the approach adopted by Good and Tideman in 1971 but not pursued <br>
further. The same view of voting theory underlies the empirical <br>
evaluations which have taken place subsequently: elections are sampled <br>
under a vague prior, and the results are assessed under an appropriate <br>
loss function. The only feature which conceals the decision-theoretic <br>
basis is the persistent use of the term 'utility' where 'loss' would be <br>
more correct.<br>
<br>
Unfortunately the constructive approach seems to be numerically <br>
intractable in cases of interest. If the number of voters was small, the <br>
observations would provide probabilistic information which could be <br>
integrated under the prior in the normal way. But as the number of <br>
voters increases, the information becomes increasingly deterministic - <br>
it degenerates to a set of equations. And therefore two cases arise. <br>
Either the equations fully determine the parameters of the voter <br>
distribution, in which case the prior almost drops out of the <br>
calculation; or the equations constrain the voter parameters to a curved <br>
manifold in which only the prior remains to be integrated.<br>
<br>
The former case was encountered by Good and Tideman, which is why their <br>
paper ended up as Bayesianism without the prior. Unfortunately their <br>
model (a single Gaussian) is too simple to be of interest, given the <br>
optimality of Condorcet methods under it.<br>
<br>
I say that the prior 'almost' drops out of the calculation because Good <br>
and Tideman's parameters have a degree of freedom which is independent <br>
of the information in the ballots. This lies in the radial distance of <br>
the three candidates from the centre of the circle whose circumference <br>
they lie on. In general we may have prior information about this <br>
distance, and it may affect the candidates' losses, so it seems a <br>
suitable case for Bayesian treatment. But Good and Tideman adopt a <br>
squared-distance loss function, and under this loss function (and this <br>
function alone, I suspect) the radial distance is immaterial to the <br>
identity of the optimal candidate. (The authors claim that the same <br>
result applies to any loss function which depends solely on distance, <br>
but I believe this to be an error.) Thus Good - of all people - made the <br>
prior drop out completely.<br>
<br>
It would be more interesting to adopt a Gaussian mixture prior as <br>
sketched above. Even then we could perform nothing more than a <br>
computational thought experiment. A truly realistic model would have to <br>
allow for an arbitrary number of Gaussians in a space of any dimension, <br>
and incorporate valence and random effects, and would need to allow for <br>
tactical voting. It's hard to imagine any useful solution being obtainable.<br>
<br>
But if voting theory is an insoluble problem in Bayesian decision <br>
theory, then any voting method we encounter must be essentially ad hoc, <br>
even if it draws on bomb-proof reasoning from another branch of <br>
mathematics. At least we have the comfort of knowing that there is a <br>
rigorous method of evaluating the solutions which are proposed to us.<br>
<br>
CJC<br>
----<br>
Election-Methods mailing list - see <a href="https://electorama.com/em" rel="noreferrer" target="_blank">https://electorama.com/em</a> for list info<br>
</blockquote></div>