nLab exchangeability

Redirected from "exchangeable random variables".
Contents

Context

Measure and probability theory

Limits and colimits

Contents

Idea

In probability theory, a (usually infinite) sequence is called exchangeable, or sometimes symmetric, if it is invariant under finite permutations. In particular:

(See below for more details.)

Exchangeable random variables and events play a prominent role in de Finetti's theorem and in the Hewitt-Savage zero-one law.

Exchangeable measures and random variables

Let XX be a measurable space, and consider the infinite product X X^{\mathbb{N}}. A finite permutation is an isomorphism σ:X X \sigma \colon X^{\mathbb{N}}\to X^{\mathbb{N}} which is a permutation of finitely many components of X X^{\mathbb{N}}. (Equivalently, it is induced by finitely many applications of the braiding.)

A probability measure pp on X X^{\mathbb{N}} is called exchangeable or symmetric if and only if it is invariant under all finite permutations. In other words, if for all finite permutations σ:X X \sigma:X^{\mathbb{N}}\to X^{\mathbb{N}}, the pushforward measure σ *p\sigma_*p is equal to pp. Even more explicitly, for every measurable subset AA of X X^{\mathbb{N}}, we have

p(σ 1(A))=p(A). p(\sigma^{-1}(A)) \;=\; p(A) .

(Note that such a probability measure is often specified by its finite marginals, via the Kolmogorov extension theorem.)

A sequence of random variables or random elements f n:ΩXf_n:\Omega\to X on a probability space (Ω,μ)(\Omega,\mu) is said to be exchangeable, or to form an exchangeable process, if the resulting joint distribution on X X^{\mathbb{N}} is an exchangeable measure.

Similar definitions can be given for a finite product X nX^n.

In categorical probability

In a Markov category, and more generally in any symmetric monoidal category, one can similarly define an exchangeable state to be a morphism IX I\to X^\mathbb{N} from the monoidal unit to an infinite tensor product which is invariant under the action of the braiding.

In other words, finite permutations form a group acting on X X^{\mathbb{N}}, and exchangeable states form a cone over the resulting diagram. De Finetti's theorem can be interpreted as saying that the limit cone is given by a space of probability measure (such as given by probability monad).

Examples

  • A sequence of iid random variables is exchangeable.

  • Somewhat on the other extreme, perfectly correlated? random variables are also exchangeable, but in general far from independent. Their joint distribution is in the form

    p(A 1×A 2×)=q(A 1A 2) p(A_1\times A_2\times\dots) \;=\; q(A_1\cap A_2\cap\dots)

    for some distribution qq on XX.

Exchangeable events and the exchangeable sigma-algebra

Similarly to measures, an event (measurable subset) AA of the infinite product X X^{\mathbb{N}} is called an exchangeable or symmetric event if for all finite permutations σ:X X \sigma:X^{\mathbb{N}}\to X^{\mathbb{N}},

σ 1(A)=A, \sigma^{-1}(A) \;=\; A ,

i.e. if it is invariant under finite permutations.

The exchangeable events of X X^{\mathbb{N}} form a sigma-algebra, called the exchangeable sigma-algebra. It is the invariant sigma-algebra for the action of finite permutations on X X^{\mathbb{N}}.

In some contexts one is, instead, interested in events which are invariant only almost surely, that is, those events AA such that for all finite permutations σ\sigma,

p(Aσ 1(A))=p(σ 1(A)A)=0. p(A \setminus \sigma^{-1}(A)) \;=\; p(\sigma^{-1}(A)\setminus A) \;=\; 0 .

These events are sometimes called almost exchangeable, almost surely exchangeable, or even just exchangeable.
They also form a sigma-algebra, also often called the exchangeable sigma-algebra.

In probability theory, very often, this ambiguity in terminology does not cause problems, since, at least for countably many random variables, it is customary to only look at processes up to almost sure equality.

In categorical probability

In categorical probability, often the exchangeable sigma-algebra can be encoded as a colimit of the action of permutations. (The same can be said about invariant sigma-algebras of more general actions.)

In Stoch and in Markov categories, the sigma-algebra of strictly (not just almost surely) invariant sets can be seen as a colimit, both in the category of Markov kernels and of measurable functions (Moss-Perrone’23).

Similarly, in Krn and in categories of couplings, the sigma-algebra of almost surely invariant sets can be seen as a colimit compatible with the dagger structure (Ensarguet-Perrone’23). For standard Borel spaces, one can, up to isomorphism, also consider strictly invariant sets.

Properties

See also

References

category: probability

Last revised on July 21, 2024 at 14:49:00. See the history of this page for a list of all contributions to it.