Feller Evolution Systems Overview

• Feller evolution systems are two-parameter operator families acting on continuous functions vanishing at infinity, ensuring positivity, mass-preservation, and Chapman–Kolmogorov consistency.
• Their infinitesimal generators are time-dependent pseudo-differential operators of Lévy–Khintchine type that enable explicit representations and characterize nonstationary behaviors.
• Markov chain approximations using stepwise Lévy increments offer stable simulation methods and underpin robust limit theorems for time-inhomogeneous stochastic models.

A Feller evolution system is a class of two-parameter operator families describing time- and space-inhomogeneous Markov processes on Banach spaces of vanishing-at-infinity continuous functions. These systems generalize the semigroup formalism of homogeneous Feller processes to settings with explicit nonstationarities in both the generator and dependence structure, providing the analytic foundation for numerous time-inhomogeneous stochastic models in probability, partial differential equations, and applications ranging from finance to population dynamics. Central to their theory are the characterization of the infinitesimal generator as a family of time-dependent pseudo-differential operators, explicit representation theorems, and the construction of stable approximations via Markov chains with stepwise Lévy increments (Böttcher, 2013, Tu, 2018).

1. Definition and Canonical Properties of Feller Evolution Systems

Let $C_0(\R^d)$ denote the Banach space of real-valued continuous functions on $\R^d$ vanishing at infinity under the supremum norm. A Feller evolution system is a two-parameter family of linear operators

$$U(s, t): C_0(\R^d) \xrightarrow{} C_0(\R^d), \quad -\infty < s \leq t < \infty$$

with the following properties:

• Positivity and mass-preservation: $U(s, t)f \geq 0$ for $f \geq 0$; $U(s, t)1 = 1$.
• Chapman–Kolmogorov property: $U(s, t) = U(s, r) U(r, t)$ for all $s \leq r \leq t$.
• Strong continuity in both parameters: For all $f \in C_0(\R^d)$,

$$\lim_{(s', t') \to (s, t)} \| U(s', t')f - U(s, t)f \|_\infty = 0$$

uniformly on compacts.

Such operator families generalize the classical one-parameter Feller semigroups encountered for homogeneous Markov processes and allow the full two-parameter time-inhomogeneous Markov transition structure (Böttcher, 2013, Tu, 2018).

2. Infinitesimal Generator Structure and Pseudo-Differential Representation

For each fixed $s$, the right generator $A^+_s$ is defined on those $f \in C_0(\R^d)$ for which

$$A^+_s f = \lim_{h \downarrow 0} \frac{U(s, s+h)f - f}{h}$$

exists in norm. Under mild domain and regularity assumptions, each $A^+_s$ is a (generally time-dependent) pseudo-differential operator of Lévy–Khintchine type: $-\,A^+_s f(x) = (2\pi)^{-d}\int_{\R^d} e^{i\<x, \xi\>} q_+(s, x, \xi) \widehat{f}(\xi)\, d\xi$ where the symbol $q_+(s, x, \xi)$ is a continuous negative-definite function in $\xi$, and admits, via the Lévy–Ito decomposition, the canonical form: $$A^+_s f(x) = b(s, x)\cdot\nabla f(x) + \frac{1}{2}\mathrm{Tr}(a(s, x) D^2 f(x)) + \int_{y \neq 0} (f(x+y) - f(x) - \nabla f(x) \cdot y \mathbf{1}_{\{ |y| < 1 \}}) \nu_{s, x}(dy)$$ where $b$ is the drift, $a$ the diffusion matrix, and $\nu_{s, x}$ the Lévy measure (Böttcher, 2013, Tu, 2018).

3. Markov Chain Approximation by Lévy-Increment Schemes

Given a Feller evolution system with symbol $q_+(s, x, \xi)$ satisfying suitable continuity and growth bounds ($|q_+| \leq C(1 + |\xi|^2)$), one constructs a sequence of discrete-time Markov chains $Z^{(n)}(k)$ with transition kernels

$\int e^{i\<y, \xi\>} \nu_{s, x, h}(dy) = \exp(-h\, q_+(s, x, \xi)), \quad h = 1/n$

so that, as $n \to \infty$, the rescaled chains $Z^{(n)}(\lfloor nt \rfloor)$ converge in law to the Feller evolution process $X_t$, with operator-norm convergence of the associated propagators $$U(s, s+t) \to (\nu_{s, \cdot, 1/n})^{\lfloor nt \rfloor}$$ (Böttcher, 2013). This construction both establishes existence and provides a practical route to simulation.

4. Space-Time Embedding and Generator Properties in the Homogenous Lift

Any $d$-dimensional Feller evolution process corresponds canonically to a $d+1$-dimensional time-homogeneous Feller process on $\mathbb{R} \times \mathbb{R}^d$, defined by

$\widetilde{X}_t = (s + t, X_{s+t})$

with transition semigroup

$$T(t)f(s, x) = \mathbb{E}[f(s+t, X_{s+t}) \mid X_s = x]$$

and infinitesimal generator

$$\mathcal{L} f(s, x) = \frac{\partial}{\partial s} f(s, x) + A^+_s(f(s, \cdot))(x)$$

which, in pseudo-differential form, yields a (possibly discontinuous) symbol $$q_{\mathcal{L}}((s, x), (\tau, \xi)) = -i \tau + q_+(s, x, \xi)$$. Notably, if the original symbol $q_+$ is only continuous in $s$, then $q_{\mathcal{L}}$ need not be smooth or continuous in the extended covariables, precluding a uniformly regular Lévy–type structure on the enlarged state space (Böttcher, 2013, Tu, 2018).

5. Applications and Dependence Structures

Feller evolution systems serve as the fundamental analytic objects for a broad array of time-inhomogeneous processes, including those driven by inhomogeneous Lévy noise (additive processes) and solutions to time-dependent parabolic PDEs with singular drifts (Kinzebulatov, 2014). Their dependence structure, notably positive dependence properties such as association, is characterized via explicit generator and Lévy kernel conditions. For stochastically monotone jump-FEPs, all forms of positive dependence (association, weak association, positive supermodular association, orthant dependence, etc.) are captured by a single support condition: all jumps at each state-time must be confined to a positive or negative orthant (Tu, 2018).

A salient case is provided by additive processes, in which the Markovian evolution is expressible in terms of independent time-inhomogeneous increments and the generator is space-homogeneous, allowing explicit characterization and facilitating the design of models with prescribed covariance or risk properties.

6. Construction, Simulation, and Limit Theorems

Feller evolution systems admit explicit construction via Chernoff-type product formulas: any family of contractions $V(s, h)$ with suitable local generator approximation yields

$$U(s, t) = \lim_{n \to \infty} \prod_{k=0}^{n-1} V(s + k (t-s)/n, (t-s)/n)$$

which, in practice, reduces to iterated application of Markov kernels corresponding to local Lévy symbols. Markov chain approximations are easily simulable since, at each discrete time, one samples a jump from the increment law determined by the symbol at the current position and time.

Through the lifted embedding into space-time, one obtains invariance principles and functional limit theorems, reducing complex time-inhomogeneous questions to the more classical time-homogeneous context on extended state spaces (Böttcher, 2013).

7. Illustrative Examples and Structural Phenomena

Simple settings illustrate the potential for nonregularity: a process in $\mathbb{R}$ with deterministic drift changing slope at a single time $s_0$ leads to a symbol $q_+(s, x, \xi)$ with a discontinuous jump at $s_0$. On the space-time lift, the resulting generator symbol acquires a discontinuity, exemplifying the general possibility of singularities in the Feller evolution framework—behavior precluded in the time-homogeneous setting (Böttcher, 2013).

• "Feller Evolution Systems: Generators and Approximation" (Böttcher, 2013)
• "Association and other forms of positive dependence for Feller evolution systems" (Tu, 2018)
• "Feller evolution families and parabolic equations with form-bounded vector fields" (Kinzebulatov, 2014)