Tutorials › Probability Theory (Graduate) › Recognizing Tail Events in Sequences

Home  /  Tutorials  /  Independence and Zero-One Laws

Independence and Zero-One Laws · Tutorial 222 of 946

Recognizing Tail Events in Sequences

Recognize when a sequence event depends only on its tail, and verify this with finite-modification arguments and measurable representations.

Intermediate 10 min read

What You'll Learn

  • Define the tail sigma-algebra of a sequence of random elements.
  • Relate invariance under finite modifications to tail measurability on sequence space.
  • Recognize limsup and eventual-property events as tail events.
  • Express convergence of a real sequence using tail coordinates.
  • Show that convergence of a series is unchanged by modifying finitely many terms.

Why Finite Changes Matter

A tail event records a property of a sequence that cannot be decided by looking at any fixed finite initial segment. One way to recognize such a property is to ask whether changing finitely many terms can change whether the event occurs.

This test is useful for events such as “\(X_n\) belongs to \(B\) infinitely often,” “\(X_n\) is eventually nonnegative,” and “the series \(\sum_{n=1}^{\infty}X_n\) converges.” Each property depends on what happens arbitrarily far along the sequence, not on any particular finite prefix.

We will make the connection precise. Measurability matters: invariance under finite changes is a pathwise statement, while membership in a tail sigma-algebra is a measurable statement. On a measurable sequence space, for measurable events, the two conditions are equivalent.

Tail Sigma-Algebras and Finite Modifications

Let \(S\) be a nonempty measurable space, and equip the sequence space \(S^{\mathbb N}\) with its product sigma-algebra. Write \(\pi_k(x)=x_k\) for the \(k\)-th coordinate of a path \(x=(x_1,x_2,\ldots)\). The tail sigma-algebra is

$$ \mathcal T=\bigcap_{m=1}^{\infty}\sigma(\pi_k:k>m). $$

Thus, an event belongs to \(\mathcal T\) precisely when it can be described using coordinates after \(m\), for every finite \(m\). The choice \(k>m\), rather than \(k\ge m\), only shifts the index convention and does not change the intersection.

Definition: Two paths \(x,y\in S^{\mathbb N}\) are finite modifications of one another if they differ in only finitely many coordinates. A measurable event \(E\subseteq S^{\mathbb N}\) is invariant under finite modifications if \(x\in E\) holds exactly when \(y\in E\), whenever \(x\) and \(y\) differ in only finitely many coordinates.

The definition requires exact invariance, not merely agreement outside a null set. If \(X_1,X_2,\ldots\) are random elements on a probability space, an event described by a measurable tail event \(E\) of their paths is \(X^{-1}(E)\), where \(X(\omega)=(X_1(\omega),X_2(\omega),\ldots)\). This event belongs to \(\bigcap_m\sigma(X_k:k>m)\).

Finite-Modification Invariance Characterizes Tail Events

Theorem: A measurable event \(E\subseteq S^{\mathbb N}\) is in \(\mathcal T\) if and only if it is invariant under finite modifications.
Proof: First suppose \(E\in\mathcal T\). Fix \(m\). The collection of events whose membership depends only on coordinates after \(m\) is a sigma-algebra containing each coordinate event generated by \(\pi_k\) for \(k>m\). Consequently, any two paths that agree after \(m\) either both belong to \(E\) or both do not. If two paths differ in only finitely many coordinates, they agree after some \(m\), so their membership in \(E\) agrees. This proves invariance. Conversely, suppose \(E\) is measurable and invariant under finite modifications. Fix \(m\), and choose any \(s_1,\ldots,s_m\in S\). Define the map \(\iota_m:S^{\mathbb N}\to S^{\mathbb N}\) by \(\iota_m(x)=(s_1,\ldots,s_m,x_{m+1},x_{m+2},\ldots)\). This map is measurable: each of its coordinates is either constant or a coordinate projection. Since \(\iota_m(x)\) differs from \(x\) in at most the first \(m\) coordinates, invariance gives \(x\in E\) if and only if \(\iota_m(x)\in E\). Hence \(E=\iota_m^{-1}(E)\). The right-hand side is measurable using only coordinates after \(m\), so \(E\in\sigma(\pi_k:k>m)\). This holds for every \(m\), and therefore \(E\in\mathcal T\). \(\square\)

The theorem gives a practical recognition principle: establish that an event is measurable, then check whether altering any finite prefix can affect its truth. The measurable qualification is essential to the statement; invariance alone is not a definition of membership in a sigma-algebra.

Worked Example: Infinitely Often Events

Worked Example: Infinitely Often Events

Let \(B\subseteq S\) be measurable, and consider the event \(E=\{x:x_n\in B\text{ for infinitely many }n\}\). By the definition of infinitely often occurrence,

$$ E=\bigcap_{N=1}^{\infty}\ \bigcup_{n\ge N}\{x:\pi_n(x)\in B\}. $$

Fix \(m\). Removing the finitely many indices \(n\le m\) cannot change whether there are infinitely many indices with \(x_n\in B\). Therefore the same event can be written as

$$ E=\bigcap_{N>m}\ \bigcup_{n\ge N}\{x:\pi_n(x)\in B\}. $$

Every event on the right uses only coordinates after \(m\), so \(E\in\sigma(\pi_k:k>m)\). Since \(m\) was arbitrary, \(E\in\mathcal T\). Equivalently, changing finitely many terms can add or remove only finitely many occurrences of membership in \(B\), and cannot change whether membership occurs infinitely often.

Eventual Properties

The same reasoning applies to properties that hold from some index onward. Suppose \(P_n\) is a measurable property of the sequence from index \(n\) onward: formally, \(P_n\in\sigma(\pi_k:k\ge n)\). The event that \(P_n\) holds eventually is

$$ E=\bigcup_{N=1}^{\infty}\ \bigcap_{n\ge N}P_n. $$

For any fixed \(m\), a witness \(N\) can be enlarged, if necessary, so that \(N>m\). Thus the same event is the union over \(N>m\) of the corresponding intersections. Each such intersection uses only coordinates after \(m\), so \(E\in\mathcal T\). In particular, any eventual condition on individual terms, such as being nonnegative or bounded by a fixed constant, is a tail event.

Worked Example: Eventually Nonnegative Terms

For real-valued coordinates, let \(P_n=\{x:x_n\ge 0\}\). The event that the terms are eventually nonnegative is

$$ E=\bigcup_{N=1}^{\infty}\ \bigcap_{n\ge N}\{x:\pi_n(x)\ge0\}. $$

For a fixed \(m\), this is also the union over \(N>m\) of the same intersections: if nonnegativity holds from some index, it holds from every larger index. The resulting expression uses only coordinates after \(m\). Hence \(E\) belongs to every tail sigma-algebra. A finite alteration may change whether the condition holds at early indices, but it cannot change whether it holds at all sufficiently large indices.

Worked Example: Convergence of a Real Sequence

Worked Example: Convergence to Zero

For a real-valued sequence, convergence to zero can be expressed using a countable set of tolerances:

$$ E=\bigcap_{r=1}^{\infty}\ \bigcup_{N=1}^{\infty}\ \bigcap_{n\ge N} \{x:|\pi_n(x)|<1/r\}. $$

This is equivalent to the usual definition: for every \(\varepsilon>0\), the terms eventually have absolute value less than \(\varepsilon\). Indeed, for any \(\varepsilon>0\), choose \(r\) with \(1/r<\varepsilon\); conversely, each \(1/r\) is a positive tolerance.

For each fixed \(m\), the index \(N\) in the union can be required to satisfy \(N>m\), since increasing \(N\) preserves the property. The displayed event can therefore be expressed using only coordinates after \(m\), and is a tail event. This also matches the finite-modification test: replacing a finite number of terms does not affect convergence to zero.

Convergence of a Series

For a real sequence, convergence of \(\sum_{k=1}^{\infty}x_k\) means convergence of its partial sums. A direct finite-modification argument is simple: if the first \(m\) terms are replaced, then for every \(n>m\) the new \(n\)-th partial sum differs from the old one by the same finite constant. Adding a constant to all sufficiently late partial sums preserves convergence. The series may have a different sum, but it converges or diverges in the same way.

Worked Example: The Series Convergence Event

Let \(X_1,X_2,\ldots\) be real-valued random variables and set \(S_p=\sum_{k=1}^{p}X_k\). By the Cauchy criterion in \(\mathbb R\), the event that the series converges is

$$ \bigcap_{r=1}^{\infty}\ \bigcup_{N=1}^{\infty}\ \bigcap_{p,q\ge N} \{|S_p-S_q|<1/r\}. $$

Fix \(m\). For a given \(r\), any \(N\) witnessing the Cauchy condition can be increased to exceed \(m\). When \(p,q\ge N>m\), the difference \(S_p-S_q\) contains only terms with indices greater than \(m\): if \(p\ge q\), then \(S_p-S_q=\sum_{k=q+1}^{p}X_k\), and if \(q\ge p\), then \(S_p-S_q=-\sum_{k=p+1}^{q}X_k\). Thus the condition for all such \(p,q\) is measurable with respect to \(\sigma(X_k:k>m)\). The full convergence event belongs to this sigma-algebra for every \(m\), and hence is a tail event.

The Cauchy representation also makes clear why convergence, not the value of the sum, is the relevant invariant property. A finite prefix can change the sum by a constant without changing whether the partial sums are Cauchy.

A General Closure Principle

Proposition: Let \(A_1,A_2,\ldots\) be events such that \(A_n\in\sigma(X_k:k\ge n)\) for every \(n\). Then both the event that \(A_n\) occurs infinitely often and the event that \(A_n\) occurs eventually always are tail events.
Proof: The infinitely-often event is \(\bigcap_{N\ge1}\bigcup_{n\ge N}A_n\). For fixed \(m\), it is unchanged if the union and intersection are restricted to \(N>m\): deleting finitely many possible occurrences does not affect infinitely-often occurrence. For \(n\ge N>m\), we have \(A_n\in\sigma(X_k:k\ge n)\subseteq\sigma(X_k:k>m)\). The restricted expression is therefore in \(\sigma(X_k:k>m)\). The eventually-always event is \(\bigcup_{N\ge1}\bigcap_{n\ge N}A_n\). Its union may also be restricted to \(N>m\), because any valid starting index can be increased. Every event in each resulting intersection belongs to \(\sigma(X_k:k>m)\), so the entire expression does too. Both conclusions hold for every \(m\), proving that both events belong to the tail sigma-algebra. \(\square\)

Using the Recognition Test Carefully

For a sequence of independent random elements, the tail sigma-algebra is the setting of Kolmogorov’s zero–one law. The results here help identify which events qualify for that law: first specify the event and its measurability, then verify that it depends only on the tail. The law concerns the probability of such an event; tail invariance itself is a separate, pathwise fact.

It is also important to distinguish convergence from the sum’s value. Convergence of a series is unchanged by modifying finitely many terms, but the sum generally changes. Likewise, “infinitely often” is unaffected by finitely many additions or deletions, whereas “at least once” need not be: a finite alteration could create or remove the only occurrence.

Check Your Understanding

Use the definition of the tail sigma-algebra and the finite-modification test to answer these questions.

  1. Why does changing finitely many terms not change whether \(X_n\in B\) infinitely often?
  2. How can the union over starting indices for an eventual property be restricted to indices larger than a fixed \(m\)?
  3. Why does convergence of a real sequence to zero have a representation using only coordinates after any fixed \(m\)?
  4. When finitely many terms of a series are altered, what changes about its partial sums for sufficiently large indices, and what remains unchanged?
  5. Why is measurability included in the theorem characterizing tail events by finite-modification invariance?

Previous

Next

Back to Tutorials
Clinical Biostats