``Drawings since hit tables in lotteries'' and a new multivariate geometric distribution (Q1305219): Difference between revisions

In number lotteries people choose \(r\) numbers out of \(s\) and win prizes if they match at least \(r-m\) numbers from the winning \(r\)-tuple. In many countries \(r=6\) and \(s=49\). For an account of the Genoese lottery from which today's number lotteries derive, the readers may consult \textit{D. R. Bellhouse} [Stat. Sci. 6, 141-148 (1991)]. A typical misconception of many lotto players is the widespread belief that certain numbers should be ``due'' if they have not come up for a long time. In order to facilitate a suitable playing strategy, which enhances this erroneous belief, there are ``drawings since hit tables'' published weekly that indicate how many drawings have taken place since each number was last selected. The author makes the usual assumptions that the weekly drawings are independent, and that each \(r\)-tuple has the same probability \((s-r)!/s!\) of being selected as the winning tuple. Let \(Y_n=(Y_{n,1},Y_{n,2},\dots,Y_{n,s})\) denote the random vector of drawings since hit numbers after \(n\) weeks for an ``\(r\) out of \(s\) lottery'' starting with the first drawing at week \(n=1\). Obviously, the sequence \((Y_n)_{n>1}\) forms a Markov chain where the initial distribution of \(Y_1\) is uniform over all \(s\)-tuples having \(r\) zeroes and \(s-r\) ones. Moreover, there are \(s!/(s-r)!\) equally likely transitions from \(Y_n\) to \(Y_{n+1}\) making \(r\) of the components of \(Y_n\) zero and increasing each of the remaining \(s-r\) components by one. The author derives the limiting distribution of \(Y_n\) as \(n\) tends to infinity. This distribution is an \(s\)-dimensional exchangeable discrete distribution having geometric marginals. As the author notes, for each fixed \(j\), the distribution of \(\max_{m=1,\dots,n}Y_{m,j}\) is the well-known distribution of the longest success run in a sequence of Bernoulli trials if the non-occurrence of the number \(j\) is considered as a success. See, e.g. the reviewer and \textit{F. S. Makri} [Fibonacci Q. 23, 338-346 (1985; Zbl 0575.60022)]. The derivation of the distribution of the ``record outstanding time'' \(R_n=\max_{j=1,\dots,s}\max_{m=1,\dots,n}Y_{m,j}\), however seems to be an open problem.