To achieve a relative accuracy of or better, the valid
domain is restricted by equations 18 and 33. That is,
(or
) and
. The former restriction can be expressed in terms of
and
by using equation 31. For large
this
becomes:
However, these only determine the relative accuracy for each part.
But for the overall process the addition of both errors must be within
bounds. Therefore, the valid domain was measued empirically (in
MATLAB). The exact boundary (where the relative error was )
is described quite well by equation 31 with
.
In practice, equation 31 is accurate in all regions where
the Z statistic is accurate (that is, where equation 18
holds). Therefore, equation 31 can be used to determine
when the domain is valid, by testing whether . Note
that a slightly lower threshold is used in practice to be slightly
conservative. Outside this region the probability is always never
less than
was tested empirically and can be confirmed by
equations 39 and 40 in the
extreme regions.
Furthermore, to speed up the calculation, it is useful to note that if
and
, then the domain is valid, whilst for
and
the probability never goes below
,
which allows it to be calculated by conventional methods. Otherwise,
for
the test involving equation 31 is used.