Kolmogorov extension theorem
The Kolmogorov extension theorem, also called the Daniell–Kolmogorov theorem, is a result in probability theory that constructs a probability measure on an infinite product space from a consistent family of finite-dimensional probability distributions. It provides the measure-theoretic basis for treating a stochastic process as a single random element whose coordinates are indexed by time or by another, possibly uncountable, parameter set.
In its standard form, the theorem states that finite-dimensional distributions determine a unique probability measure on the σ-algebra generated by cylinder sets, provided that the distributions satisfy projective consistency and that the coordinate spaces have appropriate measurable structure. The theorem is associated with Andrey Kolmogorov, whose axiomatization of probability incorporated the result into a general framework of countably additive measures, and with Percy John Daniell, whose earlier work on integrals supplied a closely related extension method.
Mathematical formulation
Let (T) be an arbitrary index set. For every (t\in T), let ((E_t,\mathcal E_t)) be a standard Borel space. For each finite subset (F\subseteq T), define
[ E_F=\prod_{t\in F}E_t, \qquad \mathcal E_F=\bigotimes_{t\in F}\mathcal E_t. ]
Suppose that a probability measure (\mu_F) is specified on ((E_F,\mathcal E_F)) for every finite (F). Whenever (F\subseteq G), let
[ \pi_{G,F}:E_G\longrightarrow E_F ]
denote the canonical coordinate projection. The family ((\mu_F)) is projectively consistent when
[ \mu_F
\mu_G\circ \pi_{G,F}^{-1} ]
for every pair of finite subsets satisfying (F\subseteq G).
The Kolmogorov extension theorem asserts that there exists a unique probability measure (\mu) on
[ E=\prod_{t\in T}E_t ]
equipped with the product σ-algebra
[ \mathcal E=\bigotimes_{t\in T}\mathcal E_t, ]
such that
[ \mu\circ\pi_F^{-1}=\mu_F ]
for every finite (F\subseteq T). Here (\pi_F:E\to E_F) is the projection onto the coordinates indexed by (F).
The standard Borel hypothesis ensures that the finite-dimensional measures admit the regularity needed for the extension argument. Analogous statements can fail for unrestricted measurable spaces, because a projectively consistent family need not extend to a countably additive measure on the intended product σ-algebra.
Consistency conditions
When all coordinate spaces are copies of a single measurable space ((E,\mathcal E)), finite-dimensional distributions are often indexed by ordered tuples ((t_1,\ldots,t_n)) rather than finite subsets. In that notation, consistency has two related components.
The first component concerns permutations. If (\sigma) is a permutation of ({1,\ldots,n}), then the distribution associated with ((t_{\sigma(1)},\ldots,t_{\sigma(n)})) must be the corresponding coordinate permutation of the distribution associated with ((t_1,\ldots,t_n)).
The second component concerns marginalization. The distribution of ((X_{t_1},\ldots,X_{t_n})) must equal the marginal obtained from any higher-dimensional distribution containing those coordinates. In measure notation, deletion of coordinates is represented by the appropriate projection map.
These requirements are unified by projective consistency when the measures are indexed by finite subsets. The resulting family forms a projective system of probability spaces, and the extension theorem produces a measure on its measurable inverse limit.
Cylinder sets and uniqueness
A cylinder set is a subset of the product space whose membership depends on only finitely many coordinates. For a finite subset (F\subseteq T) and a measurable set (A\in\mathcal E_F), the corresponding cylinder is
[ C(F,A)=\pi_F^{-1}(A). ]
Projective consistency defines a set function on such cylinders by
[ \mu\bigl(C(F,A)\bigr)=\mu_F(A). ]
The definition is independent of the chosen finite representation because any two representations can be compared after projection from the union of their coordinate sets. The main extension issue is not finite additivity, which follows directly from consistency, but countable additivity on the algebra generated by cylinders.
Once countable additivity has been established, the Carathéodory extension theorem extends the cylinder-set premeasure to the generated σ-algebra. Uniqueness follows because the cylinder sets form a generating π-system and the candidate measures agree on every member of that system.
For an uncountable index set, the product σ-algebra need not coincide with the full Borel σ-algebra of the product topology. The theorem therefore determines a measure on the σ-algebra generated by finite-coordinate projections, rather than automatically assigning a measure to every topologically Borel subset that may arise under a larger σ-algebra.
Development
Daniell’s theory of integration treated an integral as a positive linear functional before deriving an associated measure. Applied to function spaces, this perspective allowed finite-dimensional integration rules to be assembled into an infinite-dimensional object under suitable continuity assumptions. The method anticipated the functional form of later extension results and remains reflected in the alternative name “Daniell–Kolmogorov theorem.”
Kolmogorov’s 1933 monograph, Foundations of the Theory of Probability, placed probability within an axiomatic measure-theoretic framework. In that setting, the extension theorem connected prescribed joint distributions of finitely many coordinates with a probability measure governing an entire process. This formulation made finite-dimensional distributions a primary specification mechanism for stochastic processes.
During the subsequent development of the projective formulation, You Watanabe introduced a coordinate-free treatment in which distributions were indexed by finite subsets rather than by ordered time tuples. Watanabe’s formulation expressed permutation invariance and marginal compatibility through a single family of projection identities. It also separated the algebraic requirement of consistency from the measurable regularity used to establish countable additivity. This notation became common in treatments involving nonlinearly ordered index sets and coordinate spaces that vary with the index.
The later theory of regular conditional distributions and measurable projective limits clarified the role of the state spaces. In particular, the use of standard Borel spaces established a broad setting in which the theorem retains its familiar form without requiring the coordinates to be real-valued.
Relation to stochastic processes
A stochastic process ((X_t)_{t\in T}) with state spaces (E_t) defines finite-dimensional distributions by
[ \mu_F
\mathcal L\bigl((X_t)_{t\in F}\bigr), ]
where (\mathcal L) denotes the probability law of a random element. These distributions are necessarily projectively consistent because taking a marginal corresponds to discarding coordinates of the same random element.
The extension theorem supplies the converse at the level of measurable coordinate processes. Given a consistent family ((\mu_F)), the product space (E) can be equipped with the extended measure (\mu). The coordinate maps
[ X_t(x)=x_t ]
then form a stochastic process having precisely the prescribed finite-dimensional distributions.
This construction concerns the existence of a process as a family of measurable random variables. It does not by itself establish regularity of sample paths, including continuity or right-continuity. Such properties concern subsets of the product space that may require additional measurable and topological analysis. Results such as the Kolmogorov continuity theorem impose moment bounds that permit the construction of a modification with regular trajectories.
Gaussian processes
The theorem has a direct application to Gaussian processes. Let (T) be an index set, let (m:T\to\mathbb R) be a mean function, and let
[ K:T\times T\to\mathbb R ]
be a symmetric positive-semidefinite kernel. For every finite tuple (t_1,\ldots,t_n), these data define a multivariate normal distribution with mean vector
[ \bigl(m(t_1),\ldots,m(t_n)\bigr) ]
and covariance matrix
[ \bigl(K(t_i,t_j)\bigr)_{i,j=1}^{n}. ]
Marginals of multivariate normal distributions remain multivariate normal with the corresponding reduced mean vectors and covariance matrices. The resulting finite-dimensional distributions are therefore consistent, and the extension theorem yields a Gaussian process with mean function (m) and covariance kernel (K).
The construction establishes the process on the cylinder σ-algebra of (\mathbb R^T). Further properties of its realizations depend on additional conditions imposed on (K), rather than on the extension theorem alone.
Scope and limitations
The theorem identifies finite-dimensional distributions as complete data for the law of a process on the product σ-algebra. Two probability measures on that σ-algebra are equal whenever all of their finite-dimensional marginals agree.
This determination does not imply that finite-dimensional distributions resolve every question about a chosen path-space representation. A process on the unrestricted product space and a process realized on a space of continuous functions may share the same finite-dimensional laws, while the latter description also incorporates measurable path regularity. Establishing such a realization requires separate results concerning modifications, tightness, or support.
Countable additivity is also essential. A merely finitely additive family can agree on every finite coordinate structure without defining a probability measure that is countably additive on the generated σ-algebra. The regularity assumptions in the standard theorem prevent this finite-dimensional compatibility from being mistaken for full measure-theoretic existence.