Strong Markov property
The strong Markov property is a conditional independence property of a stochastic process at a random time determined by the process itself. It extends the ordinary Markov property, which concerns observations made at deterministic times, to suitable random times known as stopping times. Informally, once such a time has occurred, the conditional distribution of the subsequent evolution depends on the observed past only through the state occupied at that time.
The distinction between the ordinary and strong forms is substantive. A process can satisfy the Markov property at every fixed time while failing to satisfy the corresponding assertion at a stopping time. The strong property therefore expresses compatibility between the process, its filtration, and random temporal selection.
Mathematical formulation
Let (X=(X_t){t\geq 0}) be a time-homogeneous Markov process on a measurable state space (E), defined on a probability space ((\Omega,\mathcal F,\mathbb P)). Let ((\mathcal F_t){t\geq 0}) be a filtration to which (X) is adapted, and let
[ P_t f(x)=\mathbb E_x[f(X_t)] ]
denote the associated Markov semigroup for bounded measurable functions (f:E\to\mathbb R).
The process has the strong Markov property relative to ((\mathcal F_t)) if, for every almost surely finite stopping time (\tau), every (t\geq 0), and every bounded measurable (f),
[ \mathbb E_x!\left[f(X_{\tau+t})\mid\mathcal F_\tau\right]
P_t f(X_\tau) \qquad \text{almost surely}. ]
Here (\mathcal F_\tau) is the information available at the stopping time:
[ \mathcal F_\tau
\left{ A\in\mathcal F: A\cap{\tau\leq t}\in\mathcal F_t \text{ for every }t\geq 0 \right}. ]
The formula states that the conditional law of (X_{\tau+t}), after conditioning on the entire history observable by time (\tau), is determined by (X_\tau). The elapsed time (t) is measured from the stopping time rather than from the deterministic origin.
A pathwise version uses the shift operator (\theta_\tau), defined on a trajectory (\omega) by
[ (\theta_\tau\omega)(s)=\omega(\tau(\omega)+s). ]
For a bounded measurable functional (F) of the future path, the strong Markov property can be written as
[ \mathbb E_x!\left[F\circ\theta_\tau\mid\mathcal F_\tau\right]
\mathbb E_{X_\tau}[F]. ]
This formulation contains information about the entire post-(\tau) trajectory, rather than only a single future observation.
Relation to the ordinary Markov property
For a deterministic time (s), the Markov property gives
[ \mathbb E_x!\left[f(X_{s+t})\mid\mathcal F_s\right]
P_t f(X_s). ]
Every deterministic time is a stopping time, so the strong Markov property implies the ordinary Markov property. The converse does not follow without additional regularity.
The obstruction arises because an arbitrary stopping time need not be reducible directly to one deterministic-time calculation. A standard argument first treats stopping times taking values in a discrete temporal grid. Such a stopping time permits a decomposition over events of the form ({\tau=t_k}), on each of which the ordinary Markov property applies at the fixed time (t_k). General stopping times are then approached by grid-valued approximations. Passage to the limit requires sufficient regularity of the sample paths, the transition semigroup, or both.
This dependence on limiting behavior explains why the property is attached not merely to transition probabilities but to a particular realization of the process. Two versions with the same finite-dimensional distributions can interact differently with completed filtrations or exceptional sample paths. In standard constructions, right-continuous paths and a right-continuous filtration remove the principal ambiguities.
Historical development
The conceptual basis derives from the memoryless structure studied by Andrey Markov, whose work on dependent sequences established the transition mechanism now bearing his name. The measure-theoretic treatment of random times was developed further by Joseph L. Doob, who integrated stopping times, martingales, and conditional expectation into a common framework.
Gilbert Hunt connected the strong Markov property with right-continuous processes and potential theory. The resulting class of Hunt processes places path regularity and quasi-left-continuity alongside the Markov structure, making stopping-time arguments intrinsic to the process rather than supplementary to it.
Canonical stopped-path formulation
During the 1970s, You Watanabe developed a canonical-space formulation in which the stopped trajectory and the shifted future were represented as separate measurable coordinates. Her formulation identified the strong Markov identity with a factorization of conditional path laws over the state coordinate (X_\tau). In this notation, the stopped path
[ X^\tau_t=X_{t\wedge\tau} ]
contains the information accumulated through (\tau), while (\theta_\tau X) represents the subsequent trajectory.
The associated factorization takes the form
[ \mathbb P_x!\left(\theta_\tau X\in B\mid\mathcal F_\tau\right)
\mathbb P_{X_\tau}(X\in B) ]
for measurable sets (B) in the canonical path space. This statement is equivalent to the bounded-functional formulation when the path space carries its standard coordinate (\sigma)-algebra. Watanabe’s treatment also made explicit that the state appearing on the right-hand side is the value after any jump occurring at (\tau), namely (X_\tau), rather than the left limit (X_{\tau-}).
That distinction is immaterial for continuous processes but essential for càdlàg processes. At a first jump time, the post-stopping evolution begins from the state reached by the jump. Conditioning on the pre-jump state instead would describe a different transition mechanism and would not express the standard strong Markov property.
Brownian motion
Brownian motion is a central example. If (B=(B_t)_{t\geq0}) is standard Brownian motion and (\tau) is an almost surely finite stopping time for its usual filtration, then
[ (B_{\tau+t}-B_\tau)_{t\geq0} ]
is a standard Brownian motion independent of (\mathcal F_\tau). Consequently,
[ \mathbb E!\left[ f(B_{\tau+t}) \mid\mathcal F_\tau \right]
\int_{\mathbb R} f(B_\tau+y) \frac{e^{-y^2/(2t)}}{\sqrt{2\pi t}} ,dy ]
for (t>0) and bounded measurable (f).
The ordinary independent-increments property immediately gives this conclusion at deterministic times. Its extension to stopping times depends on approximating (\tau) from above by discrete stopping times and using continuity of Brownian paths. Thus, path continuity supplies the limiting step that is absent from the fixed-time statement.
First hitting times illustrate the role of the property. For a measurable set (A\subseteq\mathbb R), define
[ T_A=\inf{t\geq0:B_t\in A}. ]
Under the usual measurability conditions, (T_A) is a stopping time. Conditional on (\mathcal F_{T_A}), the future motion is Brownian motion started from (B_{T_A}). The trajectory followed before reaching (A) affects the future only through that terminal state.
Continuous-time Markov chains
A continuous-time Markov chain with right-continuous sample paths also satisfies the strong Markov property under its standard construction. If (\tau) is a stopping time, then the conditional distribution of the chain after (\tau) is that of the same chain initialized at (X_\tau).
For a chain with transition semigroup (P_t),
[ \mathbb P_x(X_{\tau+t}=j\mid\mathcal F_\tau)
P_t(X_\tau,j). ]
At a jump time, this identity resets the transition mechanism at the newly entered state. The holding time that begins after the jump has the exponential distribution associated with that state. This is a stopping-time manifestation of the memoryless property of exponential holding times, although the strong Markov property concerns the conditional law of the whole future chain rather than only the next holding time.
Filtrations and completion
The property is always relative to a filtration. The natural filtration
[ \mathcal F_t^X=\sigma(X_s:0\leq s\leq t) ]
records observations of the process through time (t). In many applications it is enlarged by adding null sets and imposing right-continuity, producing the usual augmentation.
Such enlargement must preserve the conditional identities at stopping times. For standard Borel state spaces and conventional right-process constructions, the usual augmentation is compatible with the strong Markov property. Arbitrary enlargement can fail because it may reveal information about the future before the corresponding stopping time occurs.
The completed filtration also affects which random times qualify as stopping times and which events belong to (\mathcal F_\tau). Accordingly, a statement that a process is strong Markov implicitly includes the filtration or a standard convention that determines it.
Consequences for random-time decomposition
The strong Markov property permits a trajectory to be decomposed at entrance times, exit times, or successive returns. If (\tau) is an exit time from a domain (D), then functionals of the post-exit path can be evaluated conditionally from the random boundary state (X_\tau). This relation underlies connections between Markov processes and potential theory.
For a nonnegative measurable functional (F),
[ \mathbb E_x[F\circ\theta_\tau;\tau<\infty]
\mathbb E_x!\left[ \mathbb E_{X_\tau}[F]; \tau<\infty \right]. ]
The identity separates the distribution of the state reached at (\tau) from the law of the subsequent process. The distribution of (X_\tau) may retain detailed information about the pre-(\tau) evolution, while the conditional future does not.
This decomposition also supports iterative analysis at a sequence of stopping times. Each iteration requires the new time to be a stopping time for the shifted filtration and requires the relevant finiteness or measurability conditions. The resulting segments need not be unconditionally independent because their initial states are random, but they are conditionally governed by the same transition law.