Reflected Diffusions in Polyhedral Domains

• Reflected diffusions in convex polyhedral domains are constrained stochastic differential equations that use oblique reflection to keep trajectories within a finite intersection of half-spaces.
• The theory leverages the Skorokhod problem and Neumann-type boundary conditions to ensure existence, uniqueness, and precise stationary distributions in applications like queueing systems and finance.
• Numerical schemes, including Euler-type discretizations and Monte Carlo methods, provide efficient convergence to invariant measures with provable sensitivity via pathwise differentiability.

Reflected diffusions in convex polyhedral domains refer to solutions of stochastic differential equations (SDEs) whose state variable is constrained to remain in a closed convex polyhedron in Euclidean space, with constraint enforced by an instantaneous “push” (typically along oblique directions) whenever the process attempts to exit the domain. These processes are the prototypical diffusion limits for multiclass queueing systems, stochastic networks, and many models in stochastic control and mathematical finance. The theory synthesizes stochastic analysis, PDEs with Neumann-type boundaries, the geometry of convex polyhedra, and advanced techniques from the theory of Skorokhod problems.

1. Model Definition and Foundational Properties

Let $D \subset \mathbb{R}^d$ be a closed convex polyhedron, represented as the intersection of finitely many halfspaces:

$$D = \bigcap_{i=1}^N \{x \in \mathbb{R}^d : \langle n_i, x \rangle \geq c_i \}$$

where $n_i$ are inward unit normals and $F_i = \{x : \langle n_i, x \rangle = c_i \}$ are the faces. The canonical reflected diffusion $Z$ in $D$ is described pathwise via the Skorokhod problem: For a given continuous semimartingale input path, find a pair $(Z, Y)$ such that

$$Z(t) = X(t) + Y(t), \quad Z(t) \in D, \forall t \ge 0,$$

where $Y$ is a process of bounded variation which increases only when $Z$ is on the boundary and such that its increments are in the cone generated by prescribed (possibly oblique) directions $$D = \bigcap_{i=1}^N \{x \in \mathbb{R}^d : \langle n_i, x \rangle \geq c_i \}$$0 with $$D = \bigcap_{i=1}^N \{x \in \mathbb{R}^d : \langle n_i, x \rangle \geq c_i \}$$1 on $$D = \bigcap_{i=1}^N \{x \in \mathbb{R}^d : \langle n_i, x \rangle \geq c_i \}$$2 (Budhiraja et al., 2012, Lipshutz et al., 2017).

Existence and uniqueness of such constrained SDEs are ensured under a geometric “complete-$$D = \bigcap_{i=1}^N \{x \in \mathbb{R}^d : \langle n_i, x \rangle \geq c_i \}$$3” condition on the collection $$D = \bigcap_{i=1}^N \{x \in \mathbb{R}^d : \langle n_i, x \rangle \geq c_i \}$$4 at the faces, as well as uniform ellipticity and Lipschitz continuity of drift $$D = \bigcap_{i=1}^N \{x \in \mathbb{R}^d : \langle n_i, x \rangle \geq c_i \}$$5 and diffusion $$D = \bigcap_{i=1}^N \{x \in \mathbb{R}^d : \langle n_i, x \rangle \geq c_i \}$$6. The Skorokhod map in this setting is globally Lipschitz on path space, a property that is crucial for strong approximations and sensitivity theory (Slominski, 2012).

2. Analytical Characterizations and the Submartingale Problem

An alternative—crucial in probabilistic analysis—is the submartingale problem formulation. For a suitable function class $$D = \bigcap_{i=1}^N \{x \in \mathbb{R}^d : \langle n_i, x \rangle \geq c_i \}$$7 (functions with compact support and gradients inward-pointing at the boundary, i.e., $$D = \bigcap_{i=1}^N \{x \in \mathbb{R}^d : \langle n_i, x \rangle \geq c_i \}$$8 on $$D = \bigcap_{i=1}^N \{x \in \mathbb{R}^d : \langle n_i, x \rangle \geq c_i \}$$9), a measure $n_i$0 on $n_i$1 is invariant for the reflected diffusion if and only if

$n_i$2

where $n_i$3 is the infinitesimal generator of the unconstrained diffusion (Kang et al., 2012, Kang et al., 2014). This provides a necessary and sufficient (martingale-problem) condition for stationarity in the convex polyhedral domain.

If $n_i$4 has a density $n_i$5 with respect to Lebesgue measure, stationarity translates to a second-order elliptic PDE with zero-flux Neumann-type boundary conditions: $n_i$6 where $n_i$7 and $n_i$8 is the adjoint operator (Kang et al., 2012, Kang et al., 2014).

3. Existence, Uniqueness, and Geometric Conditions

Existence and uniqueness of strong Markov reflected diffusions demand that, at every boundary point, the collection of active reflection directions $n_i$9 is linearly independent, and that stability conditions on the drift (for positive recurrence) are satisfied. In two dimensions, the geometric “angle” condition at corners is necessary and sufficient for uniqueness—including for domains with cusps. In the polygonal case with piecewise constant boundary data, the condition reduces to the completely-$F_i = \{x : \langle n_i, x \rangle = c_i \}$0 property of the reflection–normal matrices, as originally characterized by Dai–Williams (Costantini et al., 2022).

Local uniqueness (in neighborhoods of corners and cusps) together with a localization principle allow for global uniqueness under these conditions (Costantini et al., 2022). These properties underpin the well-posedness of both the SDE and the submartingale problem in convex polyhedral domains.

4. Numerical Schemes and Monte Carlo Approximations

Monte Carlo schemes based on Euler-type discretization with reflection are now standard for computing expectations with respect to the stationary law of reflected diffusions. The discretized process proceeds by updating the unconstrained SDE and then projecting back into $F_i = \{x : \langle n_i, x \rangle = c_i \}$1 via the one-step Skorokhod map: $F_i = \{x : \langle n_i, x \rangle = c_i \}$2 where $F_i = \{x : \langle n_i, x \rangle = c_i \}$3 is standard normal (or a bounded symmetric random vector for variance reduction). Weighted empirical averages converge almost surely to the invariant measure, with rates quantified for polynomial test functions. Lyapunov function techniques and a generalized Echeverria’s criterion are key in establishing consistency and tightness (Budhiraja et al., 2012).

Penalization approaches—replacing the hard constraint with a quadratic penalty—offer rigorous convergence rates, notably $F_i = \{x : \langle n_i, x \rangle = c_i \}$4 in $F_i = \{x : \langle n_i, x \rangle = c_i \}$5 for convex polyhedra, exploiting the global Lipschitz continuity of the Skorokhod map (Slominski, 2012, Mane, 29 Nov 2025).

For constrained sampling from Gibbs measures,

$F_i = \{x : \langle n_i, x \rangle = c_i \}$6

Langevin dynamics with reflection and its penalized, discretized variants attain polynomial convergence rates in Wasserstein distance to the true invariant law (Mane, 29 Nov 2025).

5. Pathwise Differentiability and Sensitivity Analysis

Reflected diffusions in convex polyhedral domains admit a theory of pathwise differentiability with respect to initial state, coefficients, and reflection directions. The pathwise derivative (or IPA process) is characterized as the unique solution of a constrained linear SDE—termed the derivative problem—incorporating jumps at times when the process hits nonsmooth (corner) regions: $F_i = \{x : \langle n_i, x \rangle = c_i \}$7 with $F_i = \{x : \langle n_i, x \rangle = c_i \}$8 representing the “free” perturbation and $F_i = \{x : \langle n_i, x \rangle = c_i \}$9 a process of bounded variation ensuring $Z$0 remains in the tangent space $Z$1 (Lipshutz et al., 2017, Lipshutz et al., 2016, Lipshutz et al., 2019).

This framework yields exact, representation-theoretic formulas for sensitivities (“Greeks”) of expectations, both at finite time and for stationary measures, often with lower variance than likelihood-ratio methods (Lipshutz et al., 2017, Lipshutz et al., 2019). The derivative process is crucial in stochastic optimization and robustness quantification for network and finance models.

6. Stationary Distributions and Tail Estimates

The existence of a unique invariant measure and explicit characterization of its tails are established under Foster–Lyapunov criteria, exploiting exponential Lyapunov functions constructed from the polyhedral geometry and drift direction (Sarantsev, 2015). Exponential convergence in total variation holds, and, under suitable structural assumptions, the invariant measure possesses finite exponential moments—yielding sharp exponential tail bounds.

Explicit formulas are available for the stationary distribution in the case of reflected Ornstein–Uhlenbeck processes and reflected Brownian motion in orthants, including product-form and skew-symmetric density constructions (Kang et al., 2012, Kang et al., 2014, Sarantsev, 2015).

7. Heat Kernel Bounds, Functional Inequalities, and Analytic Structure

The analytic properties of reflected Brownian motion in convex polyhedral domains are tightly linked to the underlying Dirichlet form. The associated semigroup inherits two-sided Gaussian (or sub-Gaussian) heat kernel bounds, scale-invariant Poincaré and Sobolev inequalities, and elliptic and parabolic Harnack inequalities—all as a consequence of the uniform domain structure of polyhedra (Murugan, 2023).

The Jones extension property enables $Z$2–Sobolev functions on $Z$3 to be extended to the whole Euclidean space with precise control, and the energy measure of the boundary vanishes identically—ensuring the boundary carries no Dirichlet form energy. These functional analytic results guarantee that, in this geometric context, reflected diffusions preserve the core analytic and probabilistic properties of unconstrained Brownian motion (Murugan, 2023).

---

Reflected diffusions in convex polyhedral domains thus provide a robust, geometrically explicit framework encompassing precise probabilistic, analytic, and computational structures, and their theory continues to underpin the analysis and design of complex stochastic systems.