Weak Feller Property in Stochastic Processes

• Weak Feller property is a continuity criterion for transition kernels in Markov processes, ensuring convergence of measures under weak topology.
• It underpins key results including the existence of invariant measures and ergodic behavior, critical for optimal control and numerical approximations.
• Applications span controlled stochastic systems, non-linear filters, and Dirichlet forms, offering less restrictive conditions than classical Feller properties.

The weak Feller property is a central concept in the modern theory of Markov processes, controlled stochastic systems, Dirichlet forms, and the analysis of partially observed and generalized dynamical systems. At its core, the weak Feller property provides a continuity criterion for transition kernels or semigroups, which, though weaker than the classical Feller condition, is sufficiently strong to guarantee key qualitative features such as existence of invariant measures, ergodic behavior, and robust approximation theory. This article surveys the definitional frameworks, main theoretical results, characterizations, and applications of the weak Feller property across various mathematical domains.

1. Definitions and Formal Frameworks

Classical Weak Feller Kernels

A stochastic kernel $T(\cdot|s,u)$ on a separable metric space $(\mathsf{S}, d)$ is weak Feller if for every sequence $(s_n, u_n) \to (s, u)$ in $\mathsf{S} \times \mathsf{U}$, the measures $T(\cdot|s_n, u_n)$ converge weakly to $T(\cdot|s,u)$. Equivalently, for every bounded continuous $f:\mathsf{S}\to\mathbb{R}$,

$$\int f(s')\,T(ds'|s_n, u_n) \xrightarrow[n\to\infty]{} \int f(s')\,T(ds'|s,u).$$

In the non-linear filter context, the state space is the set of probability measures $\mathsf{Z} = \mathcal{P}(\mathsf{X})$ on a Borel subset $\mathsf{X}\subset\mathbb{R}^n$, with the kernel $(\mathsf{S}, d)$0 defined analogously in terms of weak convergence on $(\mathsf{S}, d)$1 (Kara et al., 2018).

Dirichlet Forms and Hunt Processes

Given a regular semi-Dirichlet form $(\mathsf{S}, d)$2 on $(\mathsf{S}, d)$3, the weak Feller property is formulated analytically: Define $(\mathsf{S}, d)$4 as the space of bounded Borel functions vanishing “in measure” at infinity. $(\mathsf{S}, d)$5 is said to be weak Feller if:

• (Lo-diffusion): For every $(\mathsf{S}, d)$6, $(\mathsf{S}, d)$7.
• (Property P): For each compact $(\mathsf{S}, d)$8, there exists $(\mathsf{S}, d)$9, $(s_n, u_n) \to (s, u)$0, such that $(s_n, u_n) \to (s, u)$1 (BenAmor et al., 2022).

Locally Feller and Martingale Local Problems

A family of probability laws $(s_n, u_n) \to (s, u)$2 on path space $(s_n, u_n) \to (s, u)$3 is locally Feller (sometimes "weak Feller" in classical terminology) if the map $(s_n, u_n) \to (s, u)$4 is continuous in the local Skorokhod topology, the Markov property holds, and for each relatively compact open subset $(s_n, u_n) \to (s, u)$5 there is a true Feller family whose law up to exit from $(s_n, u_n) \to (s, u)$6 matches $(s_n, u_n) \to (s, u)$7. This can equivalently be formulated in terms of solvability and regularity properties for closed operators $(s_n, u_n) \to (s, u)$8 via the martingale-local problem (Gradinaru et al., 2017).

2. Main Theorems and Characterizations

Non-linear Filters

Two key results for filter kernels $(s_n, u_n) \to (s, u)$9 are established (Kara et al., 2018):

• Theorem 3.1 (Feinberg–Kara–Zgurovsky): If $\mathsf{S} \times \mathsf{U}$0 is weakly continuous in $\mathsf{S} \times \mathsf{U}$1 and $\mathsf{S} \times \mathsf{U}$2 is continuous in total variation, then $\mathsf{S} \times \mathsf{U}$3 is weak Feller.
• Theorem 3.2: If $\mathsf{S} \times \mathsf{U}$4 is continuous in total variation and $\mathsf{S} \times \mathsf{U}$5 is independent of $\mathsf{S} \times \mathsf{U}$6, then $\mathsf{S} \times \mathsf{U}$7 is weak Feller.

Key to the proof is a robust-kernel lemma which allows interchange of weak limits and uniform continuity for families of kernels or integrands, ensuring “average” continuity of Bayes-update maps.

Dirichlet Forms and Markov Semigroups

In the setting of semi-Dirichlet forms, the weak Feller property admits multiple equivalent characterizations (BenAmor et al., 2022):

1. $\mathsf{S} \times \mathsf{U}$8 for all $\mathsf{S} \times \mathsf{U}$9.
2. For every compact $T(\cdot|s_n, u_n)$0, $T(\cdot|s_n, u_n)$1 lies in $T(\cdot|s_n, u_n)$2 for each $T(\cdot|s_n, u_n)$3.
3. For every compact $T(\cdot|s_n, u_n)$4, the equilibrium potential $T(\cdot|s_n, u_n)$5 lies in $T(\cdot|s_n, u_n)$6.
4. Existence, for every compact $T(\cdot|s_n, u_n)$7, of a 1-excessive function $T(\cdot|s_n, u_n)$8 with $T(\cdot|s_n, u_n)$9.

The classical Feller property (strong continuity on $T(\cdot|s,u)$0) implies the weak Feller property, but the converse need not hold, especially in non-compact or non-symmetric contexts.

Locally Feller Processes

Equivalence theorems connect locally Feller families, solutions to martingale-local problems, and time-changed global Feller processes (Gradinaru et al., 2017). Under appropriate denseness and continuity conditions on the domain of the operator, one obtains unique locally Feller families via patching local solutions.

3. Examples and Counterexamples

Example Type	Satisfies Weak Feller?	Satisfies Feller?
State-update additive noise	Yes (if TV-continuity holds)	Sometimes
Observation additive noise	Yes (if TV-continuity holds)	Sometimes
$T(\cdot|s,u)$1 with discontinuities in $T(\cdot|s,u)$2	Not necessarily	No
Multiplication operator $T(\cdot|s,u)$3	Yes	Only if $T(\cdot|s,u)$4
Cheeger form on RCD$T(\cdot|s,u)$5 spaces	Yes	Yes
Jump-diffusions (non-symmetric)	Yes (under mild conditions)	Not necessarily

These cases illustrate that weak Feller properties are strictly weaker than classical Feller properties. Models can exhibit weak Feller behavior under relatively general assumptions, even when continuity or symmetry conditions are weakened or classical criteria fail (Kara et al., 2018, BenAmor et al., 2022).

4. Applications and Implications

The weak Feller property underpins several significant developments in stochastic analysis and control theory:

• Existence of Invariant Measures: Invariant (stationary) probability measures for Markov and filter processes are guaranteed in the presence of weak Feller property plus appropriate tightness or compactness requirements. This is foundational for ergodicity and asymptotic stability claims in filtering and controlled processes (Kara et al., 2018, BenAmor et al., 2022).
• Numerical Approximation: Weak Feller continuity of kernels enables finite-model approximation theorems, ensuring that numerical schemes for partially observed Markov decision processes (POMDPs) converge under minimal regularity (Kara et al., 2018).
• Optimal Control: In the belief-MDP formalism, the weak Feller property guarantees upper semi-continuity of value functions and existence of measurably selectable optimal policies, both for discounted and average-cost criteria, by classical dynamic programming arguments (Kara et al., 2018).
• Spectral Theory: For symmetric forms admitting weak Feller property, decomposition principles for the essential spectrum, compactness of local semigroup differences, and Persson-type theorems regarding the bottom of the essential spectrum can be established, extending classical results to irregular or singular spaces (e.g., RCD$T(\cdot|s,u)$6 spaces) (BenAmor et al., 2022).

5. Relationship to Feller and Strong Feller Properties

The classical Feller property requires that the semigroup $T(\cdot|s,u)$7 maps $T(\cdot|s,u)$8 into itself with strong continuity. By contrast, the weak Feller property operates either on $T(\cdot|s,u)$9 or on the broader set $f:\mathsf{S}\to\mathbb{R}$0. While Feller implies weak Feller, the implication is not reversible in general.

In geometric settings such as complete Riemannian manifolds with Ricci bounds, Feller and weak Feller can coincide due to kernel regularity. Non-local and non-symmetric jump-diffusions can be weak Feller without being (strong) Feller (BenAmor et al., 2022).

Further, locally Feller processes generalize the global Feller property by requiring Feller-type behavior only up to explosions or exit from compacts, thereby admitting processes not covered by global criteria—an essential extension for highly singular or locally well-behaved systems (Gradinaru et al., 2017).

6. Methodologies and Proof Structures

Typical proof strategies hinge on robust-kernel lemmas, tightness and equicontinuity arguments, and explicit construction of excessive functions or equilibrium potentials. For non-linear filters, arguments use “average” continuity of Bayes maps and the topology of weak convergence on spaces of probability measures (Kara et al., 2018). In Dirichlet-form contexts, the approach relies on controlling the behavior of semigroups on vanishing-at-infinity functions and bootstrapping local compactness to global spectral properties (BenAmor et al., 2022). For locally Feller processes, tightness and localization in the Skorokhod topology, as well as patching local martingale solutions, are central (Gradinaru et al., 2017).

7. Significance, Limitations, and Extensions

The weak Feller property, in its various formulations, delineates the frontier between processes and semigroups for which qualitative and quantitative analysis (e.g., invariant measures, optimal strategies, spectral decomposition) is tractable and robust, and those for which such structure may fail. While it is a strictly weaker requirement than classical Feller, it is sufficiently powerful to support both probabilistic and potential-theoretic frameworks. Recent advances demonstrate its critical role in analysis on singular spaces and in control of non-linear filtering processes, expanding both practical and theoretical horizons (Kara et al., 2018, BenAmor et al., 2022, Gradinaru et al., 2017).