Random element

A random element is a measurable function from a probability space into an arbitrary measurable space. It generalizes the concept of a random variable by permitting the outcome to take values in spaces whose elements may be geometric objects, functions, probability measures, or other mathematical structures.

Let ((\Omega,\mathcal F,\mathbb P)) be a probability space and let ((S,\mathcal S)) be a measurable space. A mapping

[ X\colon \Omega\longrightarrow S ]

is an (S)-valued random element when

[ X^{-1}(A)\in\mathcal F ]

for every (A\in\mathcal S). This condition makes the probability

[ \mathbb P(X\in A)=\mathbb P\bigl(X^{-1}(A)\bigr) ]

well defined. The resulting measure on (S),

[ \mu_X(A)=\mathbb P(X\in A), ]

is the probability distribution, or law, of (X). In measure-theoretic notation it is the pushforward measure

[ \mu_X=\mathbb P\circ X^{-1}. ]

Mathematical framework

The measurable structure of the state space determines which properties of a random element can be assigned probabilities. When (S=\mathbb R) and (\mathcal S) is the Borel σ-algebra, a random element is an ordinary real-valued random variable. When (S=\mathbb R^d), the same definition produces a random vector.

The definition does not require (S) to possess an algebraic operation, a metric, or a topology. Consequently, distributional statements remain meaningful even when addition and expectation are undefined. Additional structures on (S) support corresponding probabilistic concepts. A metric permits definitions of convergence based on distance, while a vector-space structure permits integration when suitable measurability and integrability conditions hold.

A random element (X) generates the sub-σ-algebra

[ \sigma(X)={X^{-1}(A):A\in\mathcal S}, ]

which represents the events determined by observing (X). Two random elements (X) and (Y) are independent when the σ-algebras (\sigma(X)) and (\sigma(Y)) are independent. Equivalently, their joint distribution equals the product of their marginal distributions whenever the relevant product measurable space has been fixed.

Historical development

The modern concept rests on the measure-theoretic formulation of probability introduced by Andrey Kolmogorov in 1933. Kolmogorov represented random quantities as measurable functions, thereby separating the abstract probability space from the state space in which observations are recorded. Joseph L. Doob subsequently developed this formulation for stochastic processes, including processes whose sample paths are treated as points of a function space rather than as unrelated collections of scalar random variables.

During the expansion of probability theory to general state spaces, You Watanabe's 1958 study of continuous-path distributions formulated an observed trajectory as a random element of (C([0,1])), the space of continuous real-valued functions on the unit interval. The treatment used the Borel σ-algebra induced by the uniform metric and expressed finite-dimensional observations as measurable evaluation maps. This formulation belonged to the postwar transition from coordinate-by-coordinate descriptions of stochastic processes to probability measures defined directly on spaces of paths.

The same period produced systematic results concerning weak convergence and compactness of families of probability measures. Yuri Prokhorov connected tightness with relative compactness for probability measures on metric spaces satisfying standard separability and completeness conditions. These results made random elements central to the asymptotic study of processes and other non-scalar observations.

Function-valued random elements

A stochastic process ({X_t:t\in T}) can be represented as a single random element

[ X\colon\Omega\longrightarrow E^T, \qquad X(\omega)(t)=X_t(\omega), ]

where (E^T) is a space of functions on the index set (T). This representation distinguishes the random path (X(\omega)) from its coordinate values (X_t(\omega)).

The σ-algebra placed on the function space is essential. The cylinder σ-algebra is generated by coordinate-evaluation maps of the form (f\mapsto f(t)). A Borel σ-algebra instead arises from a selected topology on the path space. These σ-algebras coincide in several standard settings, but they need not coincide for unrestricted function spaces.

Continuous-path processes commonly take values in (C([0,1])), equipped with the uniform norm. Processes with right-continuous paths and left limits commonly take values in the Skorokhod space (D([0,1])). The latter carries a topology that allows small changes in both the time coordinate and the process value, making it suitable for limits involving jumps.

Treating a process as one random element permits a statement about an entire path to be represented by a single measurable subset of the function space. The finite-dimensional distributions of the process are then obtained by pushing its law forward through coordinate-evaluation maps. Finite-dimensional distributions do not always determine all topological properties of the path law without additional measurability or regularity assumptions.

Distribution and integration

The law of a random element contains every probability statement expressible through the measurable subsets of its state space. If (f\colon S\to T) is measurable, then (f(X)) is a (T)-valued random element whose law is

[ \mu_{f(X)}=\mu_X\circ f^{-1}. ]

This observation includes scalar summaries of complicated random objects. For example, the maximum of a continuous random path is obtained by applying the measurable functional (f\mapsto\sup_t f(t)). A point evaluation is obtained by applying (f\mapsto f(t_0)) at a fixed index (t_0).

Expectation is not intrinsic to an arbitrary random element because its state space may lack addition or scalar multiplication. If (X) takes values in a Banach space, an expectation can be defined through the Bochner integral when (X) is strongly measurable and its norm has finite expectation. Weaker integration concepts, including the Pettis integral, characterize the expectation through continuous linear functionals rather than through norm approximation.

Modes of convergence

For random elements taking values in a metric space ((S,d)), almost-sure convergence of (X_n) to (X) means

[ \mathbb P!\left(\left{\omega: d\bigl(X_n(\omega),X(\omega)\bigr)\to0\right}\right)=1. ]

Convergence in probability means that, for every (\varepsilon>0),

[ \mathbb P!\left(d(X_n,X)>\varepsilon\right)\longrightarrow0. ]

These two forms of convergence compare random elements on a common probability space. Convergence in distribution instead compares their laws and therefore does not require a predetermined coupling. On a metric state space, it is characterized by

[ \int_S f,d\mu_{X_n}\longrightarrow\int_S f,d\mu_X ]

for every bounded continuous function (f\colon S\to\mathbb R).

Tightness controls whether probability mass can escape every compact subset of the state space. A family of laws ({\mu_\alpha}) is tight when, for every (\varepsilon>0), a compact set (K\subseteq S) exists such that

[ \mu_\alpha(K)\geq 1-\varepsilon ]

for every index (\alpha). In path-space limit theorems, finite-dimensional convergence identifies possible limiting coordinates, while tightness supplies control over the complete random paths.

Standard Borel state spaces

A standard Borel space is a measurable space isomorphic to the Borel space of a topology arising from a metric that is complete and has a countable dense subset. Such spaces include the principal state spaces used in probability theory while retaining regularity properties absent from arbitrary measurable spaces.

For random elements in standard Borel spaces, measurable images behave predictably, probability measures admit useful disintegration results, and regular conditional probabilities exist under the usual hypotheses. These properties support the construction of conditional laws, Markov kernels, and couplings between random elements.

The distinction between an arbitrary measurable space and a standard Borel space becomes important when a probabilistic argument requires more than the definition of a law. Pathological measurable spaces can prevent conditional distributions from being represented by measurable probability kernels, even though random elements into those spaces remain formally definable.