If we change the conditions of the experiment, that is, going on with our examples, we place the Geiger counter in a different place, or we vary the initial energy of the colliding particles (or we tag somehow the final state), we usually register different numbers of events in our reference class. This could just be due to statistical fluctuations. But it could (also) be due to a variation of the related physical quantity. It is then crucial, as well understood, to associate an uncertainty to the `measured' variation.
If the observed numbers are `large', things get rather easy, thanks to the Gaussian approximation of the probability distributions of interest. When, instead, the numbers are `small' the question can be quite troublesome (see, e.g., Refs. [2,3,4,5,6,7]). For example, Ref. [3] focus on the “errors on ratios of small numbers of events”, leading the readers astray: we are usually not interested in the ratios of `counts', but rather on the ratios of radioactivity levels or of production rates, and so on.
The aim of this paper is to review these
questions following consistently the rules of probability theory.
The initial, crucial point is to make a clear distinction
between the empirical observations
(the numbers of event of a given `experimentally defined class')
and the related physical quantities we are interested to infer,
although in a probabilistic way. We start playing with the Poisson
distribution in Sec. ,
referring to Appendix A for a reminder of how this distribution
is related
not only to the binomial (as well known),
but also to other important
distributions via the Poisson process, which has indeed its
roots in the Bernoulli process.
In Sec.
we show how to
use the Bayes' rule to infer Poisson
's from the observed
number of counts and then how to get the probability distribution
of their ratio
making an exact propagation of uncertainties, that is
from
and
.
Then in Sec.
we move to the
inference of intensities of the Poisson processes
(or `rates'
, in short), related to
by
, with
being the `observation time' -
it can be replaced by `integrated luminosity' or other
quantities to which the Poisson parameter
is proportional.
In the same section the `anxiety-inducing' [8]
question of the priors, assumed `flat' until
Sec.
, is finally tackled
and the conjugate priors are introduced,
showing, in particular, how to apply them in sequential
measurements of the same rate.
The technical question of getting the
probability distribution of the ratio of rates
is tackled in Sec.
.
Again, closed formulae are `luckily' obtained,
which can be extended to the more general problem
of getting the probability density function (pdf),
and its summaries, of a ratio of Gamma distributed variables.
When the game seems at the end, in Sec.
we modify the `graphical' model (indeed a visualization of the
underlying logical causal model) and restart the analysis,
this time really inferring directly
, as it will be clear.
The implications of the different models and of the
priors appearing in each of them will be analyzed with some care.
Finally, in Sec.
the same models are analyzed making use
of Markov Chain Monte Carlo (MCMC) methods, exploiting JAGS.
The purpose is twofold. First we want to cross-check the exact results
obtained in the previous section,
although the latter were limited to uniform
priors of the `top parents' of the causal model.
Second this allows not only to take into account more realistic priors,
but also to enlarge the models including efficiencies
and background, for which examples of graphical model
are provided. Another interesting question, that is
how to fit the ratio of rates as a function of another physical
question will be also addressed, showing how to modify the
causal model, but without entering into the details.
The related issue of `combining ratios' is also discussed
and it shows once more the importance of the underlying model.