To stay discovered: on tournament mean score sequences and the Bradley-Terry model (Q2145803)

The authors give new proofs of two results on mean score sequences in random tournaments. Within an \(n\)-team tournament, write \(x_i\) for the expected number of wins by team \(i\). The mean score sequence is then \((x_i,1\leq i\leq n)\). The first result, due to \textit{H. Joe} [Ann. Stat. 16, No. 2, 915--925 (1988; Zbl 0721.62067)] is within the framework of the Bradley-Terry model, in which the probability that team \(i\) beats team \(j\) in a given match is given by \(L(\lambda_i-\lambda_j)\), where \(L(\cdot)\) is the standard logistic function, and \(\lambda_i\) and \(\lambda_j\) are parameters associated with teams \(i\) and \(j\), respectively. Joe's result is that essentially all mean score sequences can be obtained from the Bradley-Terry model. More precisely, that the closure of the set of score sequences from the Bradley-Terry model in \(\mathbb{R}^n\) is the set of all score sequences. Here the authors give new, probabilistic proofs of this result. Secondly, the authors give a new, constructive proof of a result due to \textit{J. W. Moon} [Pac. J. Math. 13, 1343--1345 (1963; Zbl 0133.13303)] that an increasing sequence of \(n\) real numbers is a mean score sequence of some random tournament if and only if it is majorized by \((0,1,\ldots,n-1)\). The authors further use this as a starting point to discuss the construction of joint distributions in the representation theorem for convex ordering.