The initial motivation of this paper was didactic, i.e. how
to perform a sceptical combination of results
by MCMC using a convenient program,
after having got
a better insight of the problem by
Bayesian network (this is the name also used for the graphical models
we have encountered here).
The choice of the physics case was fortuitous, having been recently
personally interested in the charged kaon mass and having learned
thus about `apparent' disagreements between the most accurate measurements.
However, it is clear that this paper is far from attempting to
give a definite answer,
for which not only a `statistical'26but also a serious phenomenological analysis should be required.
For example in Ref. [5] there are interesting hints
on not well understood high order
corrections [33,34] and it would be
interesting to investigate if the question has been settled down
in the meanwhile and
what should be the effect on the published mass values, or whether
and how
its uncertain value should contribute to the overall uncertainty.
The result of this analysis is
where
stands for all the conditions
referred in section 6.2
(probability is always conditional
probability and hence so are also pdf's and moments of distributions).
The result seems in practical perfect agreement with the
PDG one reminded in Fig. 1.
But, first, the
estimated by sampling
is not trivial and definitely far from Gaussian (see
solid thick line of Fig. 15),
yielding e.g. the following probability intervals (not “C.L.'s”!):

Second, even if the numerical results coincide,
this agreement is just due to a compensation
of two effects in the PDG analysis which go into apposite directions:
- a weighted average of all nine individual results
(see Tab. 2),
with the final `error' scaled according to the
prescription, would have lead to
MeV, that is 13 keV lower
than that reported by the PDG [3];
- however,
the analysis
was not performed on the nine individual results of Tab. 2,
but on the six ones of Tab. 1, where
the precise result
MeV of Ref. [8]
had been `weakened' by the other three because of the
scaling prescription already applied by the authors.
For this reason the overall result went up
to
MeV, hence producing a bias of
keV, that is of the same size of the quote `error'.
The latter point is the surprising novelty of
this work, and it deserves another paper [35]
and perhaps further investigation to check if other, perhaps more
important results are affected by such a bias too.
It is a pleasure to tank Andrea Messina, Enrico Franco and Paolo Gauzzi
for discussions on the subject and comments on the manuscript.