The Natural Filtration of a Process

Information Revealed by Observations

A stochastic process is a sequence of random variables, but at a particular time we may want to restrict attention to what has actually been observed so far. The natural filtration makes this restriction precise: it records the information in the observations from the start through the current index. It does not assume that the observations are independent, Markovian, or stationary.

Let \((\Omega,\mathcal{F},\mathbb{P})\) be a probability space, let \(S\) be a finite or countable state space equipped with its power-set sigma-algebra, and let \((X_n)_{n\ge0}\) be a process with values in \(S\). For each \(n\), the history through time \(n\) is the vector \(H_n=(X_0,\ldots,X_n)\). Its natural sigma-algebra is

Here \(\sigma(H_n)\) means the smallest sigma-algebra on \(\Omega\) with respect to which \(H_n\) is measurable. Equivalently, it is generated by all events of the form \(\{X_k\in A\}\), where \(0\le k\le n\) and \(A\subseteq S\). We use the uncompleted generated sigma-algebra: events of probability zero are not automatically added unless they can be formed from the observations.

The history vector is itself a random variable taking values in the countable set \(S^{n+1}\). Its fibers, the sets \(\{\omega:H_n(\omega)=h\}\), partition \(\Omega\) after empty fibers are discarded. Because \(S^{n+1}\) is countable, \(\mathcal{F}_n^X\) consists of all unions of these fibers. Thus two outcomes with the same observed history cannot be distinguished by an event in \(\mathcal{F}_n^X\).

Why the Natural Filtration Grows

Proof. Each generator \(\{X_k\in A\}\) of \(\mathcal{F}_n^X\) has \(k\le n\), so it is also a generator of \(\mathcal{F}_{n+1}^X\). A sigma-algebra containing every generator of \(\mathcal{F}_n^X\) contains the sigma-algebra they generate; hence \(\mathcal{F}_n^X\subseteq\mathcal{F}_{n+1}^X\). Every \(X_k\) is a random variable on \((\Omega,\mathcal{F})\), so \(\{X_k\in A\}\in\mathcal{F}\) for every \(A\subseteq S\). Therefore \(\mathcal{F}_n^X\subseteq\mathcal{F}\).

For minimality, let \((\mathcal{G}_n)_{n\ge0}\) be any filtration on the same probability space such that \(X\) is adapted to it. Adapted means \(X_k\) is \(\mathcal{G}_k\)-measurable for each \(k\). Fix \(n\). For every \(k\le n\), the filtration property gives \(\mathcal{G}_k\subseteq\mathcal{G}_n\). Since \(X_k\) is \(\mathcal{G}_k\)-measurable, it is also \(\mathcal{G}_n\)-measurable. Thus \(\mathcal{G}_n\) contains every generator \(\{X_k\in A\}\) of \(\mathcal{F}_n^X\), and so \(\mathcal{F}_n^X\subseteq\mathcal{G}_n\). This holds for every \(n\), proving minimality. \(\square\)

The theorem says that the natural filtration is exactly the information needed to make the process adapted, with no additional information included. It may be smaller than the ambient sigma-algebra \(\mathcal{F}\). It may also have equal consecutive terms: if a new observation is determined by earlier ones, it need not reveal anything new.

Functions of the Observed History

Proof. The set \(S^{n+1}\) is countable and has its power-set sigma-algebra, so every function \(g:S^{n+1}\to\mathbb{R}\) is measurable. Since \(H_n\) is \(\mathcal{F}_n^X\)-measurable, the composition \(g(H_n)\) is \(\mathcal{F}_n^X\)-measurable.

For the converse, let \(Y\) be real-valued and \(\mathcal{F}_n^X\)-measurable. If two outcomes \(\omega,\omega'\) have the same history, then \(Y(\omega)=Y(\omega')\). To verify this, suppose instead that the two values differ. There is a rational number \(r\) strictly between them. Then one of \(\omega,\omega'\) belongs to \(\{Y\le r\}\) and the other does not. Since \(Y\) is measurable, \(\{Y\le r\}\in\mathcal{F}_n^X\); but every event in \(\mathcal{F}_n^X\) is a union of history fibers and cannot separate outcomes with the same history. This is a contradiction.

For each history \(h\) that occurs, define \(g(h)\) to be the common value of \(Y\) on its fiber. For histories that do not occur, set \(g(h)=0\). This defines \(g\) on all of \(S^{n+1}\), and by its definition \(Y=g(H_n)=g(X_0,\ldots,X_n)\) at every outcome. \(\square\)

This characterization gives a direct test: a quantity is measurable at time \(n\) exactly when it can be calculated from the observations through \(n\). It cannot depend on an unobserved future value or distinguish outcomes that have identical histories. The converse uses the uncompleted natural sigma-algebra and the countability of the history space; completing the sigma-algebra can add null-set distinctions.

Common Mistakes and Practical Checks

A frequent mistake is to treat \(\mathcal{F}_n^X\) as if it were automatically the full sigma-algebra \(\mathcal{F}\). In the first worked example, one might incorrectly claim that \(\{a\}\in\mathcal{F}_0^X\) because \(\{a\}\in\mathcal{F}\). The correction is to check the history fibers: the time-0 fiber containing \(a\) is \(\{a,b\}\), and an event in \(\mathcal{F}_0^X\) must contain both outcomes or neither. Thus \(\{a\}\notin\mathcal{F}_0^X\).

Another mistake is to use a future observation in a quantity claimed to be measurable now. For example, \(X_{n+1}\) is measurable with respect to \(\mathcal{F}_{n+1}^X\), but it is not necessarily measurable with respect to \(\mathcal{F}_n^X\). It is measurable at time \(n\) only in special cases, such as when its value is already determined by the history through \(n\). The right question is not whether the quantity is a random variable on the probability space, but whether its value is determined by the observations available at the specified index.

For a measurability check at time \(n\), list the distinct histories through \(n\), group outcomes with identical histories, and check whether the proposed event is a union of those groups. For a real-valued quantity, check whether it is constant within each group. These tests follow directly from the description of \(\mathcal{F}_n^X\) and apply whether or not the process has any special dependence structure.