Wednesday, July 25, 2018

A different justification for the Schulze Method

This is a different way to think about the Schulze Method, which might convince people who don't like the "beatpath" construction. 

I wrote about the Schulze method in the previous post, giving a justification that is based on the standard presentation of the method using beatpaths.


This is a different look at the same method, implemented a different way.

Suppose you first looked for a Condorcet winner: so you look at all the pairwise matchups, and find out who is preferred to who.  You make a matrix of pairwise preferences, and maybe for convenience subtract the number of voters who preferred Y to X from the number who preferred X to Y to get the margin of pairwise preferences.  (These matrices are in the example in the Wikipedia link to the Schulze method above -- if I were more detail-oriented, I'd put an example here.) 

Any row that's all positive means that candidate beats all others head-to-head: that's a Condorcet winner!  

But suppose there isn't a row that's all positive.  What do we do?  Well, first, we can eliminate any Condorcet loser: a candidate that is not preferred in any head-to-head matchup. (We'll spot a Condorcet loser in the matrix because it's got a row of margins that's all negative, or equivalently a column that's all positive.)[see footnote]

Then, since we want the closest thing to a Condorcet winner, let's "grade on a curve", and do one of the following (these are equivalent):


  • add something to all the margins: add 5 (for example) to all the margins of pairwise preferences  (add more if you haven't changed anything from negative to zero/positive, add less if you've changed too many elements) OR
  • find the negative margin that's closest to zero (-3 is closer than -10) and set it to zero.
Is there a row with no negatives now?  If so, that's our winner.  If not, our curve might've produced a definite loser (with a column that's all positive or zero), so eliminate that candidate [see footnote again] and curve some more, either adding more or zeroing out another negative margin, until you've got a row with no negatives.  It isn't a Condorcet winner, but it's the next best thing -- the candidate that started with the smallest pairwise "unpreferences".

[ Here's the footnote: ] technically, you want to eliminate every candidate not in the "Schwarz set", which is not just a Condorcet loser, but any group of candidates that loses to everyone outside the set.  (In other words, it doesn't matter if candidate D beats candidate E, if D and E both lose to A, B, and C, eliminate both of them - clearly the winner should be A, B, or C.)   But in terms of justification, this is a technical point; the idea is that you're looking for the "closest to Condorcet" winner.


Click here to return to Gnomicon home page

0 comments: