Asymptotic Reflection Coupling

• Asymptotic coupling by reflection is a method that uses reflected noise to contract the separation between stochastic processes in various settings, including Brownian motions, manifolds, and SPDEs.
• The approach relies on geometric reflection principles and one-dimensional comparison processes to ensure asymptotic optimality even when exact maximal reflection fails globally.
• Applications span kinetic Langevin dynamics, MCMC, and controlled diffusions, where hybrid strategies like synchronous coupling and gradient corrections enhance contractive behavior.

Asymptotic coupling by reflection is a family of coupling constructions in which the classical reflection mechanism is retained as the principal source of contraction, coalescence, or regularization, while exact reflection maximality is either unavailable, impossible, or replaced by an approximation that becomes effective in asymptotic regimes. In the cited literature, this theme appears in several distinct forms: pairwise near-maximal grand couplings for Brownian motions, comparison-based reflection couplings on manifolds, approximate reflected couplings for monotone SPDEs, reflected or hybrid couplings for Langevin-type dynamics and stochastic algorithms, and non-Markovian maximal couplings built from global isometries in sub-Riemannian geometry [(Li et al., 2019); (Wang, 2014); (Li et al., 2023); (Luo et al., 2024)]. The unifying idea is that reflected noise in the separation direction can be used to force the inter-particle distance toward zero, or to dominate it by a one-dimensional comparison process, even when exact pairwise maximality fails globally.

1. Classical reflection coupling and its analytical role

The basic prototype is the reflection coupling of two one-dimensional Brownian motions started at $\alpha<\beta$. It is defined by

$$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$

and then

$$\hat{X}_{\beta,t}=\alpha+\beta-\hat{X}_{\alpha,t}\quad \text{for }t<T,\qquad \hat{X}_{\beta,t}=\hat{X}_{\alpha,t}\quad \text{for }t\ge T.$$

Before meeting, the paths are mirror images across the midpoint; after meeting, they coalesce. In one dimension this coupling is maximal: among all couplings of $\mathrm{BM}(\alpha)$ and $\mathrm{BM}(\beta)$, it maximizes the probability of having coupled by time $s$, equivalently minimizing

$$\mathbf{P}\Big(\exists\,t\ge s:\hat{X}_{\alpha,t}\neq \hat{X}_{\beta,t}\Big) = \mathrm{erf}\!\left(\frac{|\alpha-\beta|}{2\sqrt{2s}}\right).$$

The relevant mechanism is that the difference process is itself a one-dimensional Brownian motion started at $|\alpha-\beta|$, so reflection realizes the shortest coupling time in stochastic order (Li et al., 2019).

On Riemannian manifolds the reflection idea is geometric rather than purely linear. Given $x,y\in M$, one chooses a minimal geodesic $Y_{xy}$ and reflects tangent vectors across the geodesic direction by

$$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$0

after parallel transport from $$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$1 to $$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$2. In the coupling by reflection of two $$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$3-diffusions, this makes the distance process $$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$4 comparable to a one-dimensional diffusion $$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$5 driven by

$$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$6

The survival probability

$$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$7

then becomes a time-dependent transportation cost, with reflection coupling providing the probabilistic input for monotonicity and total-variation comparison theorems under the Bakry–Émery $$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$8 condition (Kuwada et al., 2012).

These two settings already display the central dichotomy of the subject. In simple nondegenerate one-dimensional systems, reflection can be exactly maximal. In curved, constrained, degenerate, or many-particle settings, the same geometric principle persists, but usually only through comparison, approximation, or asymptotic optimality.

2. Pairwise near-maximality and the obstruction to exact grand reflection

The sharpest formulation of asymptotic reflection optimality in the Brownian setting is the grand-coupling problem for the family $$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$9. Reflection coupling is inherently binary: it couples one pair by mirroring one trajectory across the midpoint determined by the other. This cannot be imposed consistently on every pair in a single grand coupling. The Brownian web is the canonical coalescing grand coupling, but its pairwise coupling time is the same as twice the reflection coupling time, so it has a multiplicative gap $$\hat{X}_{\beta,t}=\alpha+\beta-\hat{X}_{\alpha,t}\quad \text{for }t<T,\qquad \hat{X}_{\beta,t}=\hat{X}_{\alpha,t}\quad \text{for }t\ge T.$$0 from the pairwise optimum, and that gap does not vanish as time tends to $$\hat{X}_{\beta,t}=\alpha+\beta-\hat{X}_{\alpha,t}\quad \text{for }t<T,\qquad \hat{X}_{\beta,t}=\hat{X}_{\alpha,t}\quad \text{for }t\ge T.$$1 or $$\hat{X}_{\beta,t}=\alpha+\beta-\hat{X}_{\alpha,t}\quad \text{for }t<T,\qquad \hat{X}_{\beta,t}=\hat{X}_{\alpha,t}\quad \text{for }t\ge T.$$2 (Li et al., 2019).

To address this, a dyadic grand coupling $$\hat{X}_{\beta,t}=\alpha+\beta-\hat{X}_{\alpha,t}\quad \text{for }t<T,\qquad \hat{X}_{\beta,t}=\hat{X}_{\alpha,t}\quad \text{for }t\ge T.$$3 is constructed using a $$\hat{X}_{\beta,t}=\alpha+\beta-\hat{X}_{\alpha,t}\quad \text{for }t<T,\qquad \hat{X}_{\beta,t}=\hat{X}_{\alpha,t}\quad \text{for }t\ge T.$$4 process $$\hat{X}_{\beta,t}=\alpha+\beta-\hat{X}_{\alpha,t}\quad \text{for }t<T,\qquad \hat{X}_{\beta,t}=\hat{X}_{\alpha,t}\quad \text{for }t\ge T.$$5, i.i.d. signs $$\hat{X}_{\beta,t}=\alpha+\beta-\hat{X}_{\alpha,t}\quad \text{for }t<T,\qquad \hat{X}_{\beta,t}=\hat{X}_{\alpha,t}\quad \text{for }t\ge T.$$6, and a random phase $$\hat{X}_{\beta,t}=\alpha+\beta-\hat{X}_{\alpha,t}\quad \text{for }t<T,\qquad \hat{X}_{\beta,t}=\hat{X}_{\alpha,t}\quad \text{for }t\ge T.$$7. Dyadic-scale signs $$\hat{X}_{\beta,t}=\alpha+\beta-\hat{X}_{\alpha,t}\quad \text{for }t<T,\qquad \hat{X}_{\beta,t}=\hat{X}_{\alpha,t}\quad \text{for }t\ge T.$$8 are assembled from the $$\hat{X}_{\beta,t}=\alpha+\beta-\hat{X}_{\alpha,t}\quad \text{for }t<T,\qquad \hat{X}_{\beta,t}=\hat{X}_{\alpha,t}\quad \text{for }t\ge T.$$9’s, and the coupled trajectories are pieced together through excursions of $\mathrm{BM}(\alpha)$0 between levels

$\mathrm{BM}(\alpha)$1

For $\mathrm{BM}(\alpha)$2, if $\mathrm{BM}(\alpha)$3,

$\mathrm{BM}(\alpha)$4

Each marginal is a Brownian motion from $\mathrm{BM}(\alpha)$5, and the result is a grand coupling of all starting points (Li et al., 2019).

The associated interval coupling time is

$\mathrm{BM}(\alpha)$6

to be compared with the pairwise reflection coupling time $\mathrm{BM}(\alpha)$7. The principal comparison is stochastic: $\mathrm{BM}(\alpha)$8 where $\mathrm{BM}(\alpha)$9 satisfies

$\mathrm{BM}(\beta)$0

Thus the grand coupling is never better than the pairwise maximal reflection coupling, but it is uniformly within a factor $\mathrm{BM}(\beta)$1 and becomes asymptotically optimal at both small and large scales. Corollary 1 gives the matching tail and quantile asymptotics: $\mathrm{BM}(\beta)$2

The same work proves that a pairwise exactly maximal grand coupling does not exist. In particular, there is no grand coupling $\mathrm{BM}(\beta)$3 such that for all $\mathrm{BM}(\beta)$4,

$\mathrm{BM}(\beta)$5

The general lower-bound inequality of Theorem 2 rules out the reflection failure function as an attainable global bound and even shows that $\mathrm{BM}(\beta)$6 is not an attainable multiplicative gap. A frequent misconception is therefore that the pairwise maximal one-dimensional reflection law should extend verbatim to all starting points; the impossibility theorem shows that global consistency constraints prevent this (Li et al., 2019).

3. Reflection couplings on manifolds, transport costs, and time-dependent geometry

For $\mathrm{BM}(\beta)$7 on a complete $\mathrm{BM}(\beta)$8-dimensional Riemannian manifold, reflection coupling under the Bakry–Émery condition

$\mathrm{BM}(\beta)$9

or, equivalently,

$s$0

produces a one-dimensional comparison distance process whose survival probability defines a concave, increasing cost function. If $s$1 are heat distributions at time $s$2 started from $s$3, the principal monotonicity statement is that for each $s$4,

$s$5

is nonincreasing on $s$6. On model spaceforms this cost is sharp because reflection coupling is maximal there and

$s$7

The resulting total-variation comparison theorem and the stability under measured Gromov–Hausdorff convergence place reflection coupling at the center of a transport-based comparison theory for heat kernels (Kuwada et al., 2012).

The time-inhomogeneous extension replaces a static metric by a family $s$8 and the generator by

$s$9

A central assumption is the lower Ricci-type bound

$$\mathbf{P}\Big(\exists\,t\ge s:\hat{X}_{\alpha,t}\neq \hat{X}_{\beta,t}\Big) = \mathrm{erf}\!\left(\frac{|\alpha-\beta|}{2\sqrt{2s}}\right).$$0

which includes backward Ricci flow through $$\mathbf{P}\Big(\exists\,t\ge s:\hat{X}_{\alpha,t}\neq \hat{X}_{\beta,t}\Big) = \mathrm{erf}\!\left(\frac{|\alpha-\beta|}{2\sqrt{2s}}\right).$$1. Rather than constructing a reflected diffusion directly, the method proceeds via time-discretized geodesic random walks on the mesh $$\mathbf{P}\Big(\exists\,t\ge s:\hat{X}_{\alpha,t}\neq \hat{X}_{\beta,t}\Big) = \mathrm{erf}\!\left(\frac{|\alpha-\beta|}{2\sqrt{2s}}\right).$$2, proves convergence

$$\mathbf{P}\Big(\exists\,t\ge s:\hat{X}_{\alpha,t}\neq \hat{X}_{\beta,t}\Big) = \mathrm{erf}\!\left(\frac{|\alpha-\beta|}{2\sqrt{2s}}\right).$$3

and derives the reflected coupling in the limit. The coupling-time estimate is expressed through a one-dimensional Ornstein–Uhlenbeck comparison process

$$\mathbf{P}\Big(\exists\,t\ge s:\hat{X}_{\alpha,t}\neq \hat{X}_{\beta,t}\Big) = \mathrm{erf}\!\left(\frac{|\alpha-\beta|}{2\sqrt{2s}}\right).$$4

with

$$\mathbf{P}\Big(\exists\,t\ge s:\hat{X}_{\alpha,t}\neq \hat{X}_{\beta,t}\Big) = \mathrm{erf}\!\left(\frac{|\alpha-\beta|}{2\sqrt{2s}}\right).$$5

This discrete-to-continuous route avoids direct cut-locus difficulties and yields gradient estimates for the semigroup as a corollary (Kuwada, 2012).

A plausible implication of these manifold results is that “asymptotic reflection” in geometry often means domination of the distance process by a one-dimensional reflected or comparison process, rather than literal pathwise reflection across a fixed hyperplane.

4. Non-Markovian reflection principles in sub-Riemannian geometry

On several sub-Riemannian manifolds with large symmetry groups, the reflection mechanism becomes non-Markovian and non-co-adapted. For the three-dimensional Heisenberg group, higher-dimensional non-isotropic Heisenberg groups, $$\mathbf{P}\Big(\exists\,t\ge s:\hat{X}_{\alpha,t}\neq \hat{X}_{\beta,t}\Big) = \mathrm{erf}\!\left(\frac{|\alpha-\beta|}{2\sqrt{2s}}\right).$$6 and its universal cover, and $$\mathbf{P}\Big(\exists\,t\ge s:\hat{X}_{\alpha,t}\neq \hat{X}_{\beta,t}\Big) = \mathrm{erf}\!\left(\frac{|\alpha-\beta|}{2\sqrt{2s}}\right).$$7, maximal couplings are obtained for Brownian motions started on the same vertical fiber by choosing a global isometry at a random stopping time, namely the first time the vertical component hits the equidistant hypersurface between the two initial points (Luo et al., 2024).

The generic structure is

$$\mathbf{P}\Big(\exists\,t\ge s:\hat{X}_{\alpha,t}\neq \hat{X}_{\beta,t}\Big) = \mathrm{erf}\!\left(\frac{|\alpha-\beta|}{2\sqrt{2s}}\right).$$8

where $$\mathbf{P}\Big(\exists\,t\ge s:\hat{X}_{\alpha,t}\neq \hat{X}_{\beta,t}\Big) = \mathrm{erf}\!\left(\frac{|\alpha-\beta|}{2\sqrt{2s}}\right).$$9 is the hitting time of the relevant hypersurface and $|\alpha-\beta|$0 is a global isometry determined by the random location $|\alpha-\beta|$1. In contrast with classical Euclidean reflection coupling, the reflecting object is not fixed in advance. Banerjee–Gordina–Mariano and Benefice had earlier constructed non-Markovian couplings in specific cases using Brownian-bridge methods; the global-isometry construction makes the maximality argument simpler and more uniform across geometries (Luo et al., 2024).

The decisive structural statement is a reflection principle for the coupling time: $|\alpha-\beta|$2 where $|\alpha-\beta|$3 is one side of the equidistant hypersurface. This identifies the coupling time exactly with a vertical hitting time and yields maximality through equality with the total variation distance. For the three-dimensional Heisenberg group, if the initial points are $|\alpha-\beta|$4 and $|\alpha-\beta|$5, then

$|\alpha-\beta|$6

so that

$|\alpha-\beta|$7

For non-isotropic Heisenberg groups the corresponding tail is of order $|\alpha-\beta|$8; for $|\alpha-\beta|$9 and its universal cover the decay is of order $x,y\in M$0; for $x,y\in M$1 the tail decays exponentially (Luo et al., 2024).

These examples sharpen an important distinction. In Euclidean one dimension, reflection is Markovian and co-adapted. In sub-Riemannian settings with vertical fibers, sharp coupling rates can require non-Markovian reflection chosen at a random geometric event. The notion of asymptotic reflection thus includes not only approximate maximality, but also reflection principles implemented through random global symmetries.

5. Infinite-dimensional equations and reflected domains

In non-linear monotone SPDEs on a Gelfand triple

$x,y\in M$2

the ideal reflection of the additive noise $x,y\in M$3 in the direction of $x,y\in M$4 is not directly available, because $x,y\in M$5 is Hilbert–Schmidt and $x,y\in M$6 is typically not a bounded operator on all of $x,y\in M$7. The exact reflected coefficient is therefore outside the usual variational existence theory. The remedy is an approximate, cutoff-based reflection coupling indexed by $x,y\in M$8, using a smooth cutoff $x,y\in M$9 and the rank-one operator $Y_{xy}$0 built from $Y_{xy}$1. The coupled system is

$Y_{xy}$2

When $Y_{xy}$3 there is no reflection; away from the diagonal, the additive noise is reflected approximately in the direction of the current difference. This asymptotic coupling by reflection yields gradient and Hölder estimates for the semigroup, as well as exponential convergence in total variation under stronger dissipativity. The framework is illustrated by stochastic generalized porous media equations, stochastic $Y_{xy}$4-Laplacian equations, and stochastic generalized fast-diffusion equations (Wang, 2014).

A different infinite-dimensional development concerns degenerate SPDEs with reflection in a constrained domain

$Y_{xy}$5

Here the coupling is by change of measures rather than literal reflected noise, but it is explicitly designed to be reflection-compatible. Two solutions $Y_{xy}$6 and $Y_{xy}$7 share the same Brownian motion, while $Y_{xy}$8 is given an additional low-mode control

$Y_{xy}$9

The reflection term in the SPDE contributes through the variational inequality

$$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$00

which implies

$$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$01

so the boundary reflection does not destroy contraction. Under the controllability assumption $$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$02 and the positivity condition $$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$03, the resulting coupling yields the asymptotic log-Harnack inequality

$$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$04

with

$$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$05

Consequences include asymptotic heat-kernel estimates, uniqueness of the invariant probability measure, asymptotic gradient estimates, the asymptotically strong Feller property, and asymptotic irreducibility; the main application is to degenerate stochastic Navier–Stokes equations with reflection (Li et al., 3 Mar 2026).

These two SPDE strands show that in infinite dimensions asymptotic reflection is often necessitated by analytic obstructions: unbounded inverse noise operators, singular drifts in $$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$06, or reflected state constraints.

6. Langevin dynamics, SGLD, and controlled diffusions

For kinetic Langevin diffusions,

$$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$07

reflection alone is insufficient because the noise acts only in velocity. The relevant contractive variable is

$$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$08

and the coupling combines reflection and synchronous components. Reflection is used when the pair is away from the contractive hyperplane $$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$09; synchronous coupling is used near that hyperplane, where the deterministic dynamics already contracts. The analysis is carried out in a tailored semimetric

$$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$10

with $$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$11 concave and constant beyond a threshold. The resulting theorem gives explicit exponential contraction

$$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$12

with rate

$$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$13

and recovers the kinetic scaling $$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$14 for rescaled double-well potentials under appropriate tuning of the friction coefficient (Eberle et al., 2017).

For SGLD, the reflection construction must handle both discretization and minibatch noise. Two copies satisfy, for $$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$15,

$$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$16

The reflected noise creates the negative $$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$17 term needed for contraction of a concave cost

$$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$18

despite local nonconvexity. Under convexity outside a compact set and sufficiently small step size,

$$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$19

and for constant step size the chain has a unique invariant measure and is geometrically ergodic in $$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$20 (Li et al., 2023).

Controlled diffusions provide another asymptotic-reflection regime. For the optimally controlled state

$$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$21

a controlled reflection coupling is built for two optimal processes by using the same feedback control in both components up to the coupling time. The distance $$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$22 satisfies

$$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$23

turning weak dissipativity into exponential contraction in Wasserstein distance. This yields existence and uniqueness of the ergodic HJB solution, uniform in time gradient and Hessian estimates for the finite-horizon value function, and turnpike estimates for value functions, optimal controls, and optimal processes (Conforti, 2022).

Across these works, the reflected component is rarely used in isolation. It is combined with synchronous coupling, Lyapunov weighting, concave distance transforms, or feedback regularity to overcome degeneracy, nonconvexity, or control dependence.

7. High-dimensional MCMC, Euler schemes, and conceptual limits

In the random walk Metropolis algorithm with Gaussian proposals,

$$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$24

high-dimensional analysis shifts attention from one-step meeting probability to contraction of the squared distance. The relevant quantity is the expected jump concordance

$$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$25

and a coupling is called asymptotically optimal if it maximizes the limiting EJC. Classical reflection,

$$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$26

works well in the spherical Gaussian case, but on eccentric Gaussian targets its gradient-direction correlation becomes

$$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$27

rather than $$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$28, so it no longer synchronizes acceptance behavior optimally. The gradient common random numbers coupling achieves $$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$29, is asymptotically optimal over product couplings in Gaussian regimes, and remains asymptotically optimal over all Markovian couplings for a broad class of product targets. The proposed gradient-corrected reflection coupling retains a reflective component to facilitate actual meeting near the diagonal while correcting reflection’s high-dimensional weakness. A common misconception is that maximal proposal overlap is the sole objective; the high-dimensional theory shows that scalable coupling requires contraction first, with exact meeting delegated to a second stage when the chains are already close (Papp et al., 2022).

For Euler-type schemes approximating non-dissipative SDEs, a different asymptotic reflection strategy is used. The exact and approximate solutions are coupled by a mixture of synchronous and reflected noise with cutoff $$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$30 and reflection matrix

$$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$31

Near the diagonal, $$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$32 and the coupling is synchronous; far from the diagonal, $$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$33 and the noise is reflected. The coupled system is

$$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$34

and yields non-asymptotic convergence in a Lyapunov-weighted quasi-Wasserstein distance,

$$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$35

For degenerate kinetic systems, the reflected variable is transformed to

$$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$36

and the metric is adapted to the hypoelliptic structure. This gives non-asymptotic bounds for Euler schemes without global dissipativity and a $$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$37-order $$\{\hat{X}_{\alpha,t}\}_{t\ge0}\sim \mathrm{BM}(\alpha),\qquad T:=\inf\{t\ge0:\hat{X}_{\alpha,t}=(\alpha+\beta)/2\},$$38-Wasserstein convergence rate for the kinetic Langevin sampler (Bao et al., 8 Dec 2025).

Taken together, these high-dimensional and numerical works identify a broad conceptual limit of raw reflection coupling. Exact or pure reflection is often either unstable near the diagonal, too weak in degenerate directions, or suboptimal for acceptance synchronization. The modern asymptotic approach is therefore hybrid: reflection is activated only in the regimes where it is contractive, and it is complemented by synchronous components, low-rank gradient corrections, cutoff functions, Lyapunov weights, or transformed coordinates. This suggests that asymptotic coupling by reflection is less a single construction than a design principle for retaining the contractive geometry of reflection under global constraints, discretization, degeneracy, or high dimensionality.