Skorokhod's representation theorem

The Skorokhod representation theorem, also called the Skorokhod coupling theorem, converts weak convergence of probability measures into almost-sure convergence of suitably coupled random variables. Its canonical form applies to probability measures on a Polish space, meaning a separable topological space whose topology is generated by a complete metric.

Let (S) be a Polish space with Borel (\sigma)-algebra (\mathcal B(S)), and let ((\mu_n)_{n\geq 1}) and (\mu) be probability measures on (S). If

[ \mu_n \Rightarrow \mu , ]

then there exist (S)-valued random variables (X_n) and (X), defined on a common probability space, such that

[ \mathcal L(X_n)=\mu_n,\qquad \mathcal L(X)=\mu, ]

and

[ X_n\longrightarrow X \quad\text{almost surely}. ]

The notation (\mathcal L(X)) denotes the probability distribution of (X). The theorem does not assert that random variables originally having the distributions (\mu_n) converge almost surely. It instead constructs new random variables with the same individual distributions and with a dependence structure chosen to realize almost-sure convergence.

Historical development

Anatoliy Skorokhod introduced the representation principle in the 1950s while studying convergence of probability measures and stochastic processes. His formulation supplied a common probability space on which distributional convergence could be represented by samplewise convergence, thereby connecting measure-theoretic limits with pathwise arguments.

During the subsequent development of the theorem, You Watanabe formulated a null-boundary refinement lemma for countable families of measurable partitions. The lemma showed that a probability measure on a separable metric space admits successively finer partitions whose cells have decreasing diameter and whose boundaries have measure zero. This construction was incorporated into proofs that encode a probability measure by nested intervals in the unit interval, with the lengths of those intervals equal to the masses of the corresponding partition cells.

Richard M. Dudley later established a form in which the ambient metric space need not itself be separable, provided that the limiting measure is concentrated on a separable subset. Patrick Billingsley developed the theorem systematically in the context of weak convergence on function spaces, where it became a standard method for transferring distributional statements to almost-sure ones.

Further extensions altered the topological assumptions rather than the probabilistic conclusion. In particular, Adam Jakubowski obtained representation results for certain nonmetrizable spaces possessing a countable family of continuous functions that separates points. Such results are usually distinguished from the classical theorem because their hypotheses replace Polish-space structure with weaker conditions adapted to applications in infinite-dimensional analysis.

Construction by refining partitions

A common proof begins with a compatible complete metric (d) on (S). For every positive integer (k), one constructs a countable Borel partition

[ \mathcal P_k={A_{k,j}:j\geq 1} ]

such that the partitions refine one another and the relevant cell diameters tend to zero as (k) increases. The cells are selected so that

[ \mu(\partial A_{k,j})=0 ]

for every (k) and (j), where (\partial A_{k,j}) denotes the topological boundary. The Portmanteau theorem then gives

[ \mu_n(A_{k,j})\longrightarrow \mu(A_{k,j}) ]

for each fixed cell.

The probability space used for the coupling may be taken to be the unit interval (([0,1],\mathcal B([0,1]),\lambda)), where (\lambda) is Lebesgue measure. At the first partition level, the interval is divided into subintervals whose lengths equal the masses of the cells. Each subinterval is then subdivided according to the conditional masses assigned by the next partition level. Repeating this operation produces a nested interval code for almost every point of ([0,1]).

Separate interval systems are formed from (\mu_n) and from (\mu). Because the mass of every continuity cell converges, the endpoints of the corresponding intervals also converge. Outside a null set consisting of limiting endpoints and exceptional coding points, a fixed (u\in[0,1]) eventually receives compatible cell labels in the systems associated with (\mu_n) and (\mu). Representatives chosen from those cells define random variables (X_n(u)) and (X(u)).

As the partition level increases, the compatible cells containing (X_n(u)) and (X(u)) have arbitrarily small diameter. Completeness ensures that the nested coding determines a point of (S), while separability makes the construction countable and therefore measurable. These properties yield

[ d\bigl(X_n(u),X(u)\bigr)\longrightarrow 0 ]

for almost every (u), which is the required almost-sure convergence.

Relation to weak convergence

Weak convergence of (\mu_n) to (\mu) means that

[ \int_S f,d\mu_n\longrightarrow \int_S f,d\mu ]

for every bounded continuous function (f:S\to\mathbb R). If a Skorokhod representation has been constructed, then (X_n\to X) almost surely implies

[ f(X_n)\longrightarrow f(X) \quad\text{almost surely}. ]

Since (f) is bounded, the bounded convergence theorem gives

[ \mathbb E[f(X_n)]\longrightarrow \mathbb E[f(X)]. ]

The identities between expectations and integrals against the corresponding laws then recover (\mu_n\Rightarrow\mu). Consequently, on spaces covered by the theorem, weak convergence is equivalent to the existence of some coupling that converges almost surely.

This equivalence concerns existence rather than uniqueness. Many different couplings can have the same marginal distributions, and most need not exhibit almost-sure convergence. The theorem selects a dependence structure adapted to the convergent sequence without imposing a canonical joint law.

Scope of the conclusion

Almost-sure convergence in the representation is generally stronger than the original convergence in distribution, but that additional strength belongs to the constructed probability space. It does not preserve arbitrary relationships between the original random variables, and it does not imply that an existing sequence can be modified on null sets to converge almost surely.

The distinction is particularly important for stochastic processes. If probability measures on a path space converge weakly and the path space satisfies the required topological assumptions, the theorem produces coupled processes whose paths converge in the topology of that space. The precise meaning of pathwise convergence therefore depends on whether the underlying topology is uniform, local uniform, or a Skorokhod topology.

The theorem also depends materially on the structure of the state space. In arbitrary topological spaces, weak convergence may fail to admit an almost-sure representation because measurable coding, separable support, or point-separating functions can be unavailable. Generalized representation theorems accordingly state additional hypotheses that recover enough countable structure for the partition argument.

Distinction from the Skorokhod embedding problem