Error bounds for exponential approximations of geometric convolutions (Q921701)

The contents of this paper are closely related to the queueing theory. Let \(\{X_ i\), \(i\geq 1\}\) be an i.i.d. sequence of nonnegative random variables and \(N_ 0\) be geometrically distributed \([P(N_ 0=k)=q^ kp\), \(k=0,1,2,...]\) and independent of \(\{X_ i\}\). Put \(Y_ 0=\sum^{N_ 0}_{1}X_ i\) and \(Y=\sum^{N}_{1}X_ i\), \(N=N_ 0+1\), which are referred to as geometrical convolutions. It is well-known that if \(X_ i\) have finite second moments, then \(Y_ 0/E Y_ 0\) and Y/E Y converge in distribution to an exponential with mean 1 as \(p\to 0.\) Define \(\mu =E X_ 1\), \(\mu_ 2=E X^ 2_ 1\), \(\gamma =\mu_ 2/2\mu^ 2\) and \(d(Y_ 0)\) as the sup norm distance between \(Y_ 0\) and an exponential distribution with mean E \(Y_ 0=q\mu /p\). The author gives the bound \(d(Y_ 0)\leq 2\gamma p\). This bound is asymptotically sharp as \(p\to 0\). Bounds are also obtained for \(d(Y_ 0+Z)\), where Z is a nonnegative random variable independent of \(Y_ 0\), for d(Y) and for \(d(Y^*)\), where \(Y^*\) is the stationary renewal distribution corresponding to Y. In Section 3, it is shown that if the distribution function F(X) of the NBUE condition \([F(t+x)\leq F(t)F(x)\) for all t,x\(\geq 0]\), then \(d(Y_ 0)\) is exactly equal to p.