The Setting
Take a random parameter and finitely many random objects , , that are independent given , the conditional law of being for a local parameter . Each is seen through an observation . A further random variable is then drawn from a law that depends on and, of , only on the observations . The question is what , and together say about the pieces : whether couples them, and whether the update of uses anything beyond and .
Conditional Laws and Kernels
A conditional law of given , for random variables and on a measure space with values in measurable spaces and , is a Markov kernel from to such that the joint law of is the law of followed by : for all measurable and . We then write .
The product kernel of Markov kernels from to , indexed by a finite set , is the Markov kernel from to that sends each point to the product of the measures: (1)✦
The observation kernel of a kernel from to through a measurable map is the kernel from to that sends to the law of the pair for drawn from : (2)✦
The disintegration of , for Markov, standard Borel and countably generated, is the Markov kernel from to that gives the law of given for drawn from : (3) where is the image of under .
The Claim
Theorem 1 (Local conditioning). Let be a finite set and an s-finite measure on , that is, a countable sum of finite measures. Let , and be random variables on with values, respectively, in a measurable space , in a product of nonempty standard Borel spaces, and in a measurable space . For each , let be a measurable map to a measurable space, a Markov kernel from to , and a measurable map to a countably generated measurable space; let be the disintegration of the observation kernel , and put and . Suppose that (4) and that, for a Markov kernel from to , (5) Then (6)✦
In the paper, the result reads: if with and , and the law of given depends only on , then , where is the law of given .
What Is Checked
The claim is stated over conditional laws under an s-finite measure , the product kernel (1), and the disintegrations (3) of the observation kernels (2), for a finite index set , nonempty standard Borel spaces and countably generated spaces .