Poisson approximations for \(r\)-scan processes (Q1198578)

The paper contains a wealth of results pertaining to asymptotical behaviour of extremal values of \(r\)-scan (and related) processes \(R_ i=R^{(r)}_ i=\sum^{i+r-1}_{k=1} X_ k\), \(i=1,\dots,n-r+1\), \(X_ i\) being positive i.i.d. r.v.'s. For suitably normalized order statistics \(R^*_ 1\leq R^*_ 2\leq\cdots\leq R^*_{n-r+1}\) one establishes either Poisson or normal asymptotics as \(n\to\infty\). This follows from the Poisson approximation to Bernoulli sums of counting r.v.'s \(N^ -_ n(a)=\sum^{n-r+1}_{i=1} W^ -_ i(a)\) \(\bigl(N^ +_ n(b)=\sum^{n-r+1}_{i=1} W^ +_ i(b)\bigr)\), where \(W^ -_ i(a)\) \((W^ +_ i(b))\) is the indicator of event \(R_ i\leq a\) \((R_ i>b)\) and \(a=a_ n\) \((b=b_ n)\) are appropriate threshold values. These Poisson approximations in terms of the total variation distance between distributions together with the convergence rates rely on the version of the Chen-Stein method proposed by Arratia, Goldstein and Gordon. The investigation was motivated by statistical problems of molecular biology connected with the analysis of the distribution of inhomogeneities in long DNA chains.