The above is the one i started reading. I assume the essence of the ideas would be best explained there. But the current lichess rating system is based on a variant or improvement on that, with apparently less-deterministic volatility model (must still be an assumption on the prior individual rating belief (or probability density of rating), that makes the emergent behaviour more to our liking or robust, like the 1500, or other invariants found so far).
Probably still no explicit bounds of the rating averages themselves either in the assumptions or in the predictions of large population limit distribution of those averages. i.e. no assumption of finite bound of the values possible. Gaussians and normal distribution of one variables have the whole real line as domain, and support. (where not zero).
Yes statistics of statistics.. integrals of integrals on a measurable set (real value, float). or Average (frequentist) of Expectation (could be Bayesian) of individual rating. i prefer to see a bunch of parameterized Gaussians flying around like gas molecules and colliding. But it is not a well mixed tank. so not a gaz or liquid... a chess server community ?
Glickman, Mark E., "Dynamic paired comparison models with stochastic variances" (2001), Journal of Applied Statistics, 28, 673-689. This paper describes the technical details of the Glicko-2 rating system. Click here for a more condensed version with a worked-out example.
http://www.glicko.net/research/dpcmsv.pdfhttp://www.glicko.net/glicko/glicko2.pdfBy the way, i think the link on lichess blog or Faq, or was it github, is broken... Here from above posted web site.