Statistical Process Monitoring with Integrated Moving Average Noise 261
The explicit eąuation for gk ( |v) in terms of <j>(') is provided in the proof of Theorem 1 in the Appendix. Upper taił areas of the distribution of U are found by integrating the density of U given in Theorem 1. Notę that gk (|v) in Theorem 3, and hence the entire distribution of U, depends on the
IMA parameter X.. Of course, it also depends on the sample size n. No large sample limiting form is known.
In Figurę 8 solid lines show the upper tails of the cumulative distribution functions (CDFs) of U for X = 0.5 and te{2,5,10,20,40}. Dashed lines show approximate CDFs derived in the next subsection. FORTRAN subroutines to calculate these distributions are available from the author. The complexity of the computation is 0(tm3) when m-point ąuadrature is used to approximate integrals. Only quadratures with evenly spaced points can be used because of the recursive naturę of Fk and Fk. To get good precision in Figurę 8, we used the trapazoidal rule with m = 161 ąuadrature points.
Approximate nuli distribution Worsley (1983) derived an excellent approximation for the upper taił of the nuli distribution of U for the case y=0. His approach uses an improved form of Bonferroni's ineąuality and it extends immediately to other values of y. The approximation is
P,{U>u}z £Pi{|Z,|> »}-£Pr{|Z,.,| > u and \Zt\> u}
The terms in the first summation are trivial to compute because Zk ~ N(0, l). The terms in the second summation come from rectangular
regions under the bivariate normal distribution of (Zk_t,Zk). (Notę that
{Zk_u Zkj is given in the proof of Theorem 1 in the Appendix.) Good
numerical functions to evaluate bivariate normal CDFs are widely available. Thus, both computational time and complexity for the approximation are minimal.
The dashed lines in Figurę 8 demonstrate that this approximation is excellent for large u. It is certainly adeąuate for gaging how much evidence there is that a shift occurred in a given set of data.