1 Definition and Formal Setup
A stopping time is a random time at which a stochastic process is “allowed” to stop, with the constraint that the decision to stop uses only information revealed by that time. This models observation rules in systems evolving randomly over time.
1.1 Filtrations and Adapted Processes
A filtration \((\mathcal{F}_t)_{t\ge 0}\) is a nondecreasing family of \(\sigma\)-algebras representing the information available up to each time \(t\). A process \(X=(X_t)\) is adapted if for every \(t\), the random variable \(X_t\) is measurable with respect to \(\mathcal{F}_t\). Adaptedness formalizes the idea that the process value at time \(t\) depends only on information through \(t\).
1.2 Measurability Criterion for Stopping Times
A random time \(\tau:\Omega\to[0,\infty]\) is a stopping time relative to \((\mathcal{F}_t)\) if for each \(t\ge 0\), \[ \{\tau \le t\}\in \mathcal{F}_t. \] Equivalently, for each time horizon \(t\), the event “the system has already stopped by time \(t\)” is determined by observations up to \(t\). This single measurability condition is the standard definition.
1.3 Random Times vs. Stopping Times
Not every random time qualifies. A random time \(\sigma\) could depend on future information (e.g., the whole future path) and thus fail the condition above. Stopping times are precisely those random times whose occurrence can be recognized using the same information stream the model uses to define \(\mathcal{F}_t\).
1.4 Examples in Discrete Time
In discrete time with \(\mathcal{F}_n\) representing information through time \(n\), a rule like \[ \tau=\inf\{n\ge 0: X_n\in A\} \] is typically a stopping time whenever \(X_n\) is adapted. Intuitively, you can check at each step whether the process has entered the set \(A\) by time \(n\).
Another example: if \(Y_n\) is \(\mathcal{F}_n\)-measurable and you define \(\tau\) as the first index \(n\) where \(Y_n=1\), then \(\{\tau\le n\}\) is determined by \(Y_0,\dots,Y_n\), hence belongs to \(\mathcal{F}_n\).
1.5 Examples in Continuous Time
In continuous time, the definition remains the same: \(\{\tau\le t\}\) must lie in \(\mathcal{F}_t\) for every \(t\). A common example is the hitting time \[ \tau=\inf\{t\ge 0: X_t\ge a\}, \] for a continuous-time adapted process \(X\). When paths have sufficient regularity, such events can be checked progressively as time increases.
2 Intuition and Interpretation
Stopping times encode a constraint on “when you’re allowed to decide.” They are central because many probabilistic results require that decisions at \(\tau\) are not secretly informed by what happens after \(\tau\).
2.1 “Stopping Based on What You Know”
The condition \(\{\tau\le t\}\in\mathcal{F}_t\) formalizes a natural rule: at time \(t\), you should be able to tell whether stopping has already occurred without seeing future randomness. The stopping rule adapts to the unfolding path.
2.2 The Information Flow Viewpoint
One can view \(\mathcal{F}_t\) as the growing record of observations. A stopping time is then an “alarm clock” whose ring time is computable from the record up to that moment. Two sample paths that share the same history up to time \(t\) must agree on whether \(\tau\le t\).
2.3 Common Stopping Rules
Typical stopping rules include:
- Boundary-based stopping: stop when a process crosses a threshold.
- Set-entry stopping: stop when the state enters a region \(A\).
- Rule at a deterministic deadline: stop at a fixed time \(T\) possibly modified by a condition on earlier data.
- After a condition becomes true: stop at the first time some adapted event holds.
2.4 Edge Cases and Pathological Behaviors
Stopping times may take the value \(\infty\), meaning “never stop.” Some constructions in continuous time require care: when filtrations are not right-continuous or processes lack regularity, certain intuitive “first time an event occurs” candidates can fail to be stopping times. These issues motivate later discussion of debut times and filtration assumptions.
3 Constructing Stopping Times
Many stopping times arise from hitting or exit events. The key is verifying that the rule respects the filtration.
3.1 First Hitting (Entry) Times
For a set \(A\) in the state space of an adapted process \(X\), the first hitting time is \[ \tau_A=\inf\{t\ge 0: X_t\in A\}. \] When \(A\) is measurable and the filtration captures the observation of \(X_t\), \(\tau_A\) is a stopping time in standard settings. Formally, the event \(\{\tau_A\le t\}\) can be expressed in terms of whether \(X_s\in A\) for some \(s\le t\), which is determined by information available up to \(t\).
3.2 Exit Times from Sets
For an open or measurable set \(D\), the exit time \[ \tau_D=\inf\{t\ge 0: X_t\notin D\} \] captures the first moment the process leaves \(D\). This is useful for confinement problems, absorption phenomena, and boundary crossing analyses. Under typical regularity, the exit event by time \(t\) depends only on the path up to \(t\).
3.3 Last Exit Times
A last exit time considers the final time the process is in a set: \[ L_A=\sup\{t\ge 0: X_t\in A\}, \] with the convention that the supremum of an empty set is \(0\) or \(-\infty\) depending on context. Last-exit times are subtler than first-hit times: whether they are stopping times depends strongly on how the event “no later re-entry occurs” is encoded in the filtration, since that depends on future behavior relative to the putative time.
3.4 Regime-Switching and Threshold Rules
Stopping times can be designed to respond to the system’s regime. For example, define \(\tau\) as the first time a Markov or adapted process enters a “high-risk” region, or as the first time a statistic exceeds a threshold. Such rules often reduce to first-hitting times of a derived process (e.g., an accumulated loss or log-likelihood ratio).
3.5 Stopping at a Prescribed Time
Deterministic stopping is a special case: \(\tau=T\) is a stopping time for any constant \(T\). More generally, one can combine conditions with a deterministic deadline, such as stopping at \[ \tau = T\wedge \tau_A, \] which stops at time \(T\) if the set entry has not occurred earlier.
4 Operations and Closure Properties
Stopping times form a rich algebra under natural operations. These closure properties help build complex stopping strategies from simpler ones.
4.1 Minimum and Maximum of Stopping Times
If \(\tau\) and \(\sigma\) are stopping times, then \(\tau\wedge\sigma\) (the earlier one) and \(\tau\vee\sigma\) (the later one) are also stopping times. The events \(\{\tau\wedge\sigma\le t\}\) and \(\{\tau\vee\sigma\le t\}\) can be expressed using \(\{\tau\le t\}\) and \(\{\sigma\le t\}\), ensuring measurability with respect to \(\mathcal{F}_t\).
4.2 Sums with Deterministic Times
If \(\tau\) is a stopping time and \(c\ge 0\) is deterministic, then \(\tau+c\) is a stopping time relative to the shifted-time filtration notion appropriate to the model. In the usual convention with the same filtration indexed by time, measurability follows from \[ \{\tau+c\le t\}=\{\tau\le t-c\} \] for \(t\ge c\), and the appropriate empty-event behavior for \(t<c\).
4.3 Stopping at a Stopping Time (Nested Stopping)
A central construction is to stop a process at a random time that itself is determined by the filtration. If you define a time like “the first time after \(\sigma\) that a condition holds,” then the resulting random time is often expressible as a function of \(\sigma\) and a shifted rule. This idea is formalized in the strong Markov property context and in nested stopping constructions.
4.4 Limits of Stopping Times
If \((\tau_n)\) is an increasing sequence of stopping times, then the limit \(\tau=\lim_{n\to\infty}\tau_n\) is a stopping time. Similarly, if \(\tau_n\) decreases, the limit (with suitable conventions) is also a stopping time. These results come from the stability of measurability under countable unions and intersections.
4.5 Measurable Transformations of Stopping Times
If \(\tau\) is a stopping time and \(f:[0,\infty]\to[0,\infty]\) is Borel measurable and nondecreasing, then \(f(\tau)\) is a stopping time. Monotonicity ensures that events of the form \(\{f(\tau)\le t\}\) can be rewritten in terms of \(\{\tau\le s\}\) for some \(s\).
5 Stopped Processes
Stopping times naturally lead to stopped processes, which freeze the evolution of a process after the stopping time occurs.
5.1 Definition of a Stopped Process
Given an adapted process \(X=(X_t)\) and a stopping time \(\tau\), the stopped process \(X^\tau\) is typically defined by \[ X^\tau_t = X_{t\wedge\tau}. \] Thus, until time \(\tau\) the process evolves normally, and after \(\tau\) it remains at its value at the stopping moment.
5.2 Adaptedness and Stopping-Time Measurability
If \(X\) is adapted, then \(X^\tau\) is adapted as well. For each \(t\), the value \(X^\tau_t\) equals \(X_t\) when \(t\le \tau\) and equals \(X_\tau\) when \(\tau\le t\). In standard frameworks, the measurability of \(\{t\le\tau\}\) with respect to \(\mathcal{F}_t\) supports this adaptedness.
5.3 Sample-Path Interpretation
On a single sample path, stopping replaces the original trajectory \((X_t(\omega))_{t\ge 0}\) with a truncated-and-held version: the path is followed up to \(\tau(\omega)\), and then kept constant. This provides an intuitive bridge between probabilistic definitions and pathwise behavior.
5.4 Continuity Considerations
If \(X\) has continuous paths, then \(X^\tau\) inherits continuity (assuming stopping at a finite time along the same continuous trajectory). If \(X\) has jumps, \(X^\tau\) preserves the relevant jump structure up to the stopping time and then freezes thereafter, which can affect martingale properties unless integrability conditions are met.
5.5 Stopped Martingales
A fundamental fact is that under appropriate conditions, a stopped martingale remains a martingale or becomes a martingale-like object. For example, for integrable martingales \(X\), the stopped process \(X^\tau\) is often a martingale when \(\tau\) is bounded or satisfies integrability constraints. This is the basis for “optional stopping” arguments.
6 Optional Stopping and Martingale Applications
Optional stopping results connect the value of a martingale at stopping times to its initial expectation. These theorems are among the most used tools in probability.
6.1 Optional Stopping Theorem (Statement and Conditions)
A typical form states: if \(X\) is a martingale and \(\tau\) is a stopping time meeting certain conditions (commonly boundedness or integrability), then \[ \mathbb{E}[X_\tau]=\mathbb{E}[X_0]. \] Different versions of the theorem specify alternative sufficient conditions. The key role of measurability ensures that \(X_\tau\) is not biased by future information.
6.2 Uniform Integrability Approach
One robust condition is uniform integrability of the stopped family. If \((X_{t\wedge\tau})_{t\ge 0}\) is uniformly integrable, then optional stopping can be justified even when \(\tau\) is unbounded. This approach avoids some failures that occur with weaker integrability assumptions.
6.3 Bounded Stopping Times
If \(\tau\) is almost surely bounded (e.g., \(\tau\le T\)), then optional stopping often holds under minimal assumptions. Boundedness limits the time horizon, reducing technical issues related to convergence and heavy tails.
6.4 Almost-Sure Termination and Integrability
| If the process stops almost surely but \(\mathbb{E}[ | X_\tau | ]\) is not controlled, optional stopping may fail. Thus, alongside almost-sure termination, one typically needs integrability conditions ensuring that expectations and limits can be interchanged safely. |
|---|
6.5 Applications to Fair Games
Martingale optional stopping is used to analyze “fair games” where expected gains remain zero under the rules of the model. By choosing stopping rules that depend only on current information—such as stopping upon reaching a win threshold—one can compute expected outcomes and justify strategies within the model’s assumptions.
7 Stopping Times in Key Models
Stopping times appear in many canonical stochastic processes, providing a language for boundary crossing, transaction triggers, and observation endpoints.
7.1 Random Walks and Hitting Boundaries
In a one-dimensional random walk, stopping times often represent the first time the walk hits an upper or lower boundary. These times are used to study ruin probabilities, expected durations until absorption, and fair-play stopping strategies.
A common setup is: let \(S_n\) be partial sums of independent increments, and stop at the first \(n\) where \(S_n\) reaches either a profit level or a loss level. The stopping time is a first hitting time of a discrete set.
7.2 Brownian Motion Stopping Rules
For Brownian motion \(B_t\), stopping times frequently involve hitting levels or exiting intervals. Many results leverage martingales derived from Brownian motion (e.g., exponential martingales or harmonic functions), combined with optional stopping under conditions suited to continuous-time behavior.
Examples include:
- stopping when \(B_t\) first hits a specified barrier,
- stopping at the exit time from an interval \((a,b)\).
7.3 Poisson Process Arrival-Based Stopping
For a Poisson process \(N_t\), stopping times can be defined by event counts, such as the time of the \(k\)-th arrival: \[ \tau_k=\inf\{t\ge 0: N_t=k\}. \] These \(\tau_k\) are stopping times relative to the filtration generated by the Poisson process. They are fundamental in queueing, reliability, and renewal theory.
7.4 Markov Processes and Strong Markov Property
For Markov processes, stopping times that fit the model’s natural filtration allow the use of the strong Markov property, which states that after a stopping time, the process behaves like a fresh Markov process started from the state at the stopping moment. This turns many global questions into local ones and supports recursive analyses.
7.5 Queueing and Renewal-Type Stopping
In queueing systems and renewal models, stopping times represent milestones such as:
- time until a system becomes empty,
- time until the \(k\)-th service completion,
- time until a workload exceeds a threshold.
Because these quantities are determined by the observed evolution of arrival and service events up to the time in question, they fit naturally into the stopping-time framework used to evaluate performance metrics.
8 Further Topics and Extensions
Stopping times extend beyond basic settings into refined distinctions about predictability, filtration structure, and optimization.
8.1 Predictable vs. Optional Times
Not all stopping times are equally “accessible.” A predictable time can be announced by an increasing sequence of earlier stopping times. Predictability matters in stochastic integration and in the study of compensators and jump processes, where optional and predictable times interact differently with martingale decompositions.
8.2 Debut Times and Right-Continuous Filtrations
Debut times describe the earliest occurrence of an adapted set-valued event. Under right-continuity assumptions on the filtration and suitable measurability, debut times provide stopping times for many “first occurrence” constructions. These results address the subtle failures that can occur when filtrations are not well-behaved.
8.3 Stopping Times Relative to Enlarged Information
One may consider a filtration larger than the one originally used, representing access to additional information (e.g., delayed signals or auxiliary randomness). A time that is a stopping time under one filtration may fail under another. Studying such changes clarifies how the stopping rule’s legitimacy depends on the information structure.
8.4 Optimal Stopping Problem (Connections)
The optimal stopping problem asks for a stopping time that maximizes or minimizes an expected payoff, often written as \(\mathbb{E}[G_\tau]\). This connects stopping times with dynamic programming principles, variational inequalities, and free-boundary problems in diffusion settings. Here, the concept of “admissible” stopping times is determined by the filtration: only information available up to \(\tau\) may be used.
8.5 Stopping in Stochastic Control Contexts
In stochastic control, one often optimizes a combination of control actions and stopping decisions. Stopping times serve as the moments when control is terminated or switched, such as stopping a process when costs exceed a limit or when a system reaches a target state. The theory blends stopping-time measurability with control-theoretic Hamilton–Jacobi–Bellman methods and martingale techniques.