摘要：We thank the two discussant for their insightful comments. The starting point of
this work is described in the third paragraph of Section 2.4 in the manuscript and
consists of obtaining a likelihood for using the distribution of the quadratic form
(z ..())0..1(z ..()). Thus represents the covariance of the di
erence z ..(),
that we denote . The common approach in the optimal ngerprint literature is to
estimate from a long control run, that we denote wj ; j = 1; : : : ; k and plug it into
the likelihood. One of our key interests in developing this work was that of calibrating
with a method that accounted for the estimation uncertainties involved in estimating
both and the parameters needed to obtain a surrogate model for . Thus, from
the onset, our goal was not to `cut feedback', borrowing the expression from Rougier's
discussion. This is in contrast with the more traditional approach that consists of
tting a Gaussian process to the model output, obtain estimates for the parameters
that describe such process and plug them into the likelihood used for the calibration
of . Something that is lost by not cutting feedback is the ability to evaluate the
emulator o
-line. On the other hand, we feel that too often developing and checking
the emulator has received too much emphasis. In a calibration problem the focus is
on exploring the parameter values that produce the most plausible computer model
responses when compared to actual observations. When, as is the case in our problem,
an abundant set of computer evaluations is available, it is likely that the precision of
the emulator will be of secondary importance
