11.5.2.1. Defining the percentile

The first method discussed here is straightforward: the principle of rejection is to keep the best-tested configurations. This can be done either by applying a threshold value on the distance results (called \(\delta\) in [Bla17]) or by retaining a fixed fraction of the tested configurations, defined through a percentile \(\varepsilon_{Dist}\). The TRejectionABC method implements the latter approach. By default, the percentile is set to \(\varepsilon_{Dist} = 0.01\).

To modify this value, the user can call setPercentile, whose prototype is

void setPercentile(double eps);

where the argument eps specifies the fraction of configurations to be kept.

An important consequence is that the total number of configurations evaluated is computed as follows:

\[n_{\rm Comp} = \frac{n_{S}}{\varepsilon_{Dist}}\]

where \(n_{S}\) is the number of retained samples in the final posterior distribution, as defined in the constructor (see Constructing the TRejectionABC object).