Let be a sequence of measurable spaces, each equipped with two measures Pn and Qn.
We say that Qn is contiguous with respect to Pn (denoted Qn ◁ Pn) if for every sequence An of measurable sets, Pn(An) → 0 implies Qn(An) → 0.
The sequences Pn and Qn are said to be mutually contiguous or bi-contiguous (denoted Qn ◁▷ Pn) if both Qn is contiguous with respect to Pn and Pn is contiguous with respect to Qn.[2]
The notion of contiguity is closely related to that of absolute continuity. We say that a measure Q is absolutely continuous with respect to P (denoted Q ≪ P) if for any measurable set A, P(A) = 0 implies Q(A) = 0. That is, Q is absolutely continuous with respect to P if the support of Q is a subset of the support of P, except in cases where this is false, including, e.g., a measure that concentrates on an open set, because its support is a closed set and it assigns measure zero to the boundary, and so another measure may concentrate on the boundary and thus have support contained within the support of the first measure, but they will be mutually singular. In summary, this previous sentence's statement of absolute continuity is false. The contiguity property replaces this requirement with an asymptotic one: Qn is contiguous with respect to Pn if the "limiting support" of Qn is a subset of the limiting support of Pn. By the aforementioned logic, this statement is also false.
It is possible however that each of the measures Qn be absolutely continuous with respect to Pn, while the sequence Qn not being contiguous with respect to Pn.
The fundamental Radon–Nikodym theorem for absolutely continuous measures states that if Q is absolutely continuous with respect to P, then Q has density with respect to P, denoted as ƒ = dQ⁄dP, such that for any measurable set A
which is interpreted as being able to "reconstruct" the measure Q from knowing the measure P and the derivative ƒ. A similar result exists for contiguous sequences of measures, and is given by the Le Cam's third lemma.
Propertiesedit
For the case for all n it applies .
It is possible that is true for all n without .[3]
Le Cam's first lemmaedit
For two sequences of measures on measurable spaces the following statements are equivalent:[4]
for any statistics .
where and are random variables on .
Interpretationedit
Prohorov's theorem tells us that given a sequence of probability measures, every subsequence has a further subsequence which converges weakly. Le Cam's first lemma shows that the properties of the associated limit points determine whether contiguity applies or not. This can be understood in analogy with the non-asymptotic notion of absolute continuity of measures.[5]
^Wolfowitz J.(1974) Review of the book: "Contiguity of Probability Measures: Some Applications in Statistics. by George G. Roussas",
Journal of the American Statistical Association, 69, 278–279 jstor
Necessary and sufficient conditions for contiguity and entire asymptotic separation of probability measures R Sh Liptser et al 1982 Russ. Math. Surv. 37 107–136
The unconscious as infinite sets By Ignacio Matte Blanco, Eric (FRW) Rayner
"Contiguity of Probability Measures", David J. Scott, La Trobe University