From the paper: Bayesian analysis is an important method of modern statistics, it almost appears in all important areas of statistical research and has been very useful in many fields of applications. A main different point between Bayesian analysis and some classical statistical methods is that we use not only the sample information but also some information about the parameter \(\theta\) in Bayesian analysis. Essential in the Baysian approach is to view the parameter \(\theta\) as a value of some random variable \(\overline\Theta\) with a known distribution (rather than viewing \(\theta\) as an unknown constant). This completely specified (discrete or continuous) density over the parameter space \(\Theta\) is called the prior density, which reflects past experience about the parameter \(\theta\). Usually, given the states of a random variable \(X\), a conditional probability is attached to this variable, say \(f(x\,|\,\theta)\), and a prior density of the parameter \(\theta\), say \(p(\theta)\), is specified based on previous knowledge.
In the present paper, we propose a criterion which tells us how to choose a loss function in Bayesian analysis.