Sticky Diffusion Processes

• Sticky diffusion is a class of continuous Markov processes that spend positive Lebesgue time at boundaries, characterized by atoms in speed measures and local time identities.
• The models employ time change formulations, Dirichlet forms, and Wentzell–Robin boundary conditions to rigorously describe behavior in one and higher dimensions.
• Applications include simulation, statistical inference, and practical implementations in finance, reaction models, and transport in crowded media.

Searching arXiv for recent and foundational work on sticky diffusion to ground the article in published sources. Sticky diffusion denotes a class of continuous Markov processes that evolve as ordinary diffusions away from a distinguished threshold, boundary, or collision set, but spend positive Lebesgue time on that lower-dimensional set once they reach it. In one dimension, this feature is classically encoded either by a speed measure with an atom at the sticky point or by a coupled local-time/occupation-time identity; in higher dimensions it is represented through Dirichlet forms with boundary mass, Wentzell–Robin boundary conditions, or time changes of nonsticky diffusions [(Anagnostakis, 2022); (Grothaus et al., 2014); (Berry et al., 2024)]. The concept has been developed for bounded Euclidean domains, hypersurfaces, metric graphs, particle-collision manifolds, and semipermeable interface problems, and it supports a substantial literature on simulation, inference, and applications in finance, reaction models, and transport in crowded media (Meier et al., 2021, Anagnostakis, 2023).

1. Defining structure and one-dimensional formulations

The defining property of stickiness is positive occupation time at a lower-dimensional set. For sticky Brownian motion on a threshold at $0$, the process behaves like Brownian motion away from $0$, but when it hits $0$ it spends a positive amount of dwell-time there proportional to a stickiness parameter. In the notation of one-dimensional sticky Itô diffusions, this is expressed by

$$dX_t = \mu(X_t)\mathbf 1_{\{X_t\neq 0\}}dt + \sigma(X_t)\mathbf 1_{\{X_t\neq 0\}}dB_t, \qquad \int_0^t \mathbf 1_{\{X_s=0\}}\,ds = \frac{\rho}{2}L_t^0(X),$$

with $L_t^0(X)$ the local time at $0$ (Anagnostakis, 2022). In the sticky Brownian special case, the scale function is $s(x)=x$ and the speed measure is $m(dx)=2\,dx+\rho\,\delta_0(dx)$, so the atomic part of the speed measure is the analytic signature of stickiness (Anagnostakis, 2022).

A broader threshold model combines stickiness with skewness and oscillating volatility. The sticky–oscillating–skew threshold diffusion is written as

$\begin{cases} X_t = X_0 + \int_0^t b(X_s)\mathbf 1_{\{X_s\neq0\}}\,ds + \int_0^t \sigma(X_s)\mathbf 1_{\{X_s\neq0\}}\,dW_s + \beta L_t^0(X),\[0.5ex] \int_0^t \mathbf 1_{\{X_s=0\}}\,ds = \frac{\rho}{2}L_t^0(X), \end{cases}$

where $\beta\in[-1,1]$ controls skewness and $0$0 controls stickiness (Anagnostakis et al., 2024). This formulation makes explicit that stickiness, skew reflection, and discontinuous diffusivity can be coupled within a single one-dimensional semimartingale model.

A standard equivalent construction is by time change. On the half-line, root-$0$1 sticky Brownian motion can be obtained from reflecting Brownian motion $0$2 with boundary local time $0$3 via

$0$4

and its generator $0$5 is supplemented by the boundary condition $0$6 (Bou-Rabee et al., 2019). On star graphs, the same Itô–McKean mechanism reappears: if $0$7 is the nonsticky diffusion and $0$8, then $0$9, with $0$0 the right-inverse of $0$1, is the sticky diffusion with vertex parameter $0$2 (Berry et al., 2024).

These representations distinguish sticky diffusion from ordinary reflection. For reflected Brownian motion, the boundary local time measures repeated contact, but the process does not accumulate positive Lebesgue time at the boundary. In sticky models, by contrast, the occupation identity $0$3 or its higher-dimensional analog is part of the definition (Anagnostakis, 2022).

Setting	Signature of stickiness	Representative source
One-dimensional threshold	Atom in speed measure; $0$4	(Anagnostakis, 2022)
Half-line Brownian model	Boundary condition $0$5; time change by local time	(Bou-Rabee et al., 2019)
Star graph vertex	$0$6	(Berry et al., 2024)

2. Dirichlet-form construction and Wentzell–Robin boundary dynamics

A general higher-dimensional construction is given for a bounded domain $0$7 with $0$8-boundary $0$9, state space $$dX_t = \mu(X_t)\mathbf 1_{\{X_t\neq 0\}}dt + \sigma(X_t)\mathbf 1_{\{X_t\neq 0\}}dB_t, \qquad \int_0^t \mathbf 1_{\{X_s=0\}}\,ds = \frac{\rho}{2}L_t^0(X),$$0, interior density $$dX_t = \mu(X_t)\mathbf 1_{\{X_t\neq 0\}}dt + \sigma(X_t)\mathbf 1_{\{X_t\neq 0\}}dB_t, \qquad \int_0^t \mathbf 1_{\{X_s=0\}}\,ds = \frac{\rho}{2}L_t^0(X),$$1, boundary density $$dX_t = \mu(X_t)\mathbf 1_{\{X_t\neq 0\}}dt + \sigma(X_t)\mathbf 1_{\{X_t\neq 0\}}dB_t, \qquad \int_0^t \mathbf 1_{\{X_s=0\}}\,ds = \frac{\rho}{2}L_t^0(X),$$2, and reference measure

$$dX_t = \mu(X_t)\mathbf 1_{\{X_t\neq 0\}}dt + \sigma(X_t)\mathbf 1_{\{X_t\neq 0\}}dB_t, \qquad \int_0^t \mathbf 1_{\{X_s=0\}}\,ds = \frac{\rho}{2}L_t^0(X),$$3

On $$dX_t = \mu(X_t)\mathbf 1_{\{X_t\neq 0\}}dt + \sigma(X_t)\mathbf 1_{\{X_t\neq 0\}}dB_t, \qquad \int_0^t \mathbf 1_{\{X_s=0\}}\,ds = \frac{\rho}{2}L_t^0(X),$$4, one considers

$$dX_t = \mu(X_t)\mathbf 1_{\{X_t\neq 0\}}dt + \sigma(X_t)\mathbf 1_{\{X_t\neq 0\}}dB_t, \qquad \int_0^t \mathbf 1_{\{X_s=0\}}\,ds = \frac{\rho}{2}L_t^0(X),$$5

with $$dX_t = \mu(X_t)\mathbf 1_{\{X_t\neq 0\}}dt + \sigma(X_t)\mathbf 1_{\{X_t\neq 0\}}dB_t, \qquad \int_0^t \mathbf 1_{\{X_s=0\}}\,ds = \frac{\rho}{2}L_t^0(X),$$6; $$dX_t = \mu(X_t)\mathbf 1_{\{X_t\neq 0\}}dt + \sigma(X_t)\mathbf 1_{\{X_t\neq 0\}}dB_t, \qquad \int_0^t \mathbf 1_{\{X_s=0\}}\,ds = \frac{\rho}{2}L_t^0(X),$$7 permits diffusion along the boundary (Grothaus et al., 2014). Under the Hamza condition, the closure is a recurrent, strongly local, regular Dirichlet form, hence generates a diffusion on $$dX_t = \mu(X_t)\mathbf 1_{\{X_t\neq 0\}}dt + \sigma(X_t)\mathbf 1_{\{X_t\neq 0\}}dB_t, \qquad \int_0^t \mathbf 1_{\{X_s=0\}}\,ds = \frac{\rho}{2}L_t^0(X),$$8 (Grothaus et al., 2014).

For $$dX_t = \mu(X_t)\mathbf 1_{\{X_t\neq 0\}}dt + \sigma(X_t)\mathbf 1_{\{X_t\neq 0\}}dB_t, \qquad \int_0^t \mathbf 1_{\{X_s=0\}}\,ds = \frac{\rho}{2}L_t^0(X),$$9, the associated generator is

$L_t^0(X)$0

or, equivalently, $L_t^0(X)$1 with $L_t^0(X)$2 and $L_t^0(X)$3 the projection onto the tangent space $L_t^0(X)$4 (Grothaus et al., 2014). The boundary condition induced in the heat equation $L_t^0(X)$5 is the general Wentzell, or Robin–Wentzell, condition

$L_t^0(X)$6

(Grothaus et al., 2014).

The same process admits an SDE formulation. In the interior,

$L_t^0(X)$7

whereas on the boundary,

$L_t^0(X)$8

Using the Revuz correspondence for the boundary local time $L_t^0(X)$9, the normal drift may be rewritten as $0$0, and this term is explicitly identified as the source of sticky behavior: after hitting $0$1, the process can remain on $0$2 for positive occupation time before being pushed back into $0$3 (Grothaus et al., 2014).

Under the additional condition $0$4, each connected component $0$5 of $0$6 is invariant, the restricted Dirichlet form is irreducible recurrent, and ergodicity holds: $0$7 for quasi every $0$8 (Grothaus et al., 2014). In particular,

$0$9

which confirms genuine stickiness rather than transient boundary contact (Grothaus et al., 2014). Under local integrability assumptions on $s(x)=x$0 and $s(x)=x$1, the process is $s(x)=x$2-strong Feller and uniquely well posed for every starting point outside $s(x)=x$3 (Grothaus et al., 2014).

3. Graph, interface, and interacting-particle generalizations

Sticky diffusion extends naturally from Euclidean boundaries to singular state spaces. On finite metric graphs, each edge carries a one-dimensional diffusion, while vertices are equipped with sticky and semipermeable transmission conditions. For the edgewise generator

$s(x)=x$4

the domain is defined by continuity of the vertex value together with sticky-flux balance conditions such as

$s(x)=x$5

and the analogous right-endpoint condition (Gregosiewicz, 2022). The resulting operator generates a conservative Feller semigroup. When the diffusion speed scales as $s(x)=x$6 and $s(x)=x$7, the fast interior motion collapses onto an effective slow jump dynamics on the graph structure, with semigroup convergence $s(x)=x$8 (Gregosiewicz, 2022).

On star graphs, the sticky effect is concentrated at the central vertex $s(x)=x$9. If the edgewise generator is

$m(dx)=2\,dx+\rho\,\delta_0(dx)$0

then the domain condition at the vertex is

$m(dx)=2\,dx+\rho\,\delta_0(dx)$1

with $m(dx)=2\,dx+\rho\,\delta_0(dx)$2 the stickiness parameter and $m(dx)=2\,dx+\rho\,\delta_0(dx)$3 the edge-selection weights (Berry et al., 2024). For $m(dx)=2\,dx+\rho\,\delta_0(dx)$4, the local time $m(dx)=2\,dx+\rho\,\delta_0(dx)$5 at the vertex satisfies

$m(dx)=2\,dx+\rho\,\delta_0(dx)$6

which is the graph analogue of the one-dimensional occupation identity (Berry et al., 2024). The paper further proves a Freidlin–Sheu type Itô formula for $m(dx)=2\,dx+\rho\,\delta_0(dx)$7, with an explicit local-time boundary term involving $m(dx)=2\,dx+\rho\,\delta_0(dx)$8 and $m(dx)=2\,dx+\rho\,\delta_0(dx)$9 (Berry et al., 2024).

A different generalization concerns collision manifolds. In the $\begin{cases} X_t = X_0 + \int_0^t b(X_s)\mathbf 1_{\{X_s\neq0\}}\,ds + \int_0^t \sigma(X_s)\mathbf 1_{\{X_s\neq0\}}\,dW_s + \beta L_t^0(X),\[0.5ex] \int_0^t \mathbf 1_{\{X_s=0\}}\,ds = \frac{\rho}{2}L_t^0(X), \end{cases}$0-particle sticky Brownian motion on $\begin{cases} X_t = X_0 + \int_0^t b(X_s)\mathbf 1_{\{X_s\neq0\}}\,ds + \int_0^t \sigma(X_s)\mathbf 1_{\{X_s\neq0\}}\,dW_s + \beta L_t^0(X),\[0.5ex] \int_0^t \mathbf 1_{\{X_s=0\}}\,ds = \frac{\rho}{2}L_t^0(X), \end{cases}$1, one works on the principal Weyl chamber $\begin{cases} X_t = X_0 + \int_0^t b(X_s)\mathbf 1_{\{X_s\neq0\}}\,ds + \int_0^t \sigma(X_s)\mathbf 1_{\{X_s\neq0\}}\,dW_s + \beta L_t^0(X),\[0.5ex] \int_0^t \mathbf 1_{\{X_s=0\}}\,ds = \frac{\rho}{2}L_t^0(X), \end{cases}$2, with generator $\begin{cases} X_t = X_0 + \int_0^t b(X_s)\mathbf 1_{\{X_s\neq0\}}\,ds + \int_0^t \sigma(X_s)\mathbf 1_{\{X_s\neq0\}}\,dW_s + \beta L_t^0(X),\[0.5ex] \int_0^t \mathbf 1_{\{X_s=0\}}\,ds = \frac{\rho}{2}L_t^0(X), \end{cases}$3 in the interior and sticky boundary conditions on each collision hyperplane $\begin{cases} X_t = X_0 + \int_0^t b(X_s)\mathbf 1_{\{X_s\neq0\}}\,ds + \int_0^t \sigma(X_s)\mathbf 1_{\{X_s\neq0\}}\,dW_s + \beta L_t^0(X),\[0.5ex] \int_0^t \mathbf 1_{\{X_s=0\}}\,ds = \frac{\rho}{2}L_t^0(X), \end{cases}$4. In the exactly solvable uniform-characteristic case, the boundary condition is

$\begin{cases} X_t = X_0 + \int_0^t b(X_s)\mathbf 1_{\{X_s\neq0\}}\,ds + \int_0^t \sigma(X_s)\mathbf 1_{\{X_s\neq0\}}\,dW_s + \beta L_t^0(X),\[0.5ex] \int_0^t \mathbf 1_{\{X_s=0\}}\,ds = \frac{\rho}{2}L_t^0(X), \end{cases}$5

and the transition density is obtained by a Bethe-Ansatz integral formula with two-body scattering amplitude

$\begin{cases} X_t = X_0 + \int_0^t b(X_s)\mathbf 1_{\{X_s\neq0\}}\,ds + \int_0^t \sigma(X_s)\mathbf 1_{\{X_s\neq0\}}\,dW_s + \beta L_t^0(X),\[0.5ex] \int_0^t \mathbf 1_{\{X_s=0\}}\,ds = \frac{\rho}{2}L_t^0(X), \end{cases}$6

(Brockington et al., 2021). This shows that sticky interaction can be imposed not only at external boundaries but also at pairwise coincidence sets.

A further interface-based development studies multivariate SDEs with countably many local-time terms on semipermeable hyperplanes. After a random time change depending on interface local times, the interfaces become sticky. In the regime $\begin{cases} X_t = X_0 + \int_0^t b(X_s)\mathbf 1_{\{X_s\neq0\}}\,ds + \int_0^t \sigma(X_s)\mathbf 1_{\{X_s\neq0\}}\,dW_s + \beta L_t^0(X),\[0.5ex] \int_0^t \mathbf 1_{\{X_s=0\}}\,ds = \frac{\rho}{2}L_t^0(X), \end{cases}$7 and $\begin{cases} X_t = X_0 + \int_0^t b(X_s)\mathbf 1_{\{X_s\neq0\}}\,ds + \int_0^t \sigma(X_s)\mathbf 1_{\{X_s\neq0\}}\,dW_s + \beta L_t^0(X),\[0.5ex] \int_0^t \mathbf 1_{\{X_s=0\}}\,ds = \frac{\rho}{2}L_t^0(X), \end{cases}$8, the local-time terms homogenize into a limiting drift, and the clock process converges to

$\begin{cases} X_t = X_0 + \int_0^t b(X_s)\mathbf 1_{\{X_s\neq0\}}\,ds + \int_0^t \sigma(X_s)\mathbf 1_{\{X_s\neq0\}}\,dW_s + \beta L_t^0(X),\[0.5ex] \int_0^t \mathbf 1_{\{X_s=0\}}\,ds = \frac{\rho}{2}L_t^0(X), \end{cases}$9

The effective coefficients are divided by $\beta\in[-1,1]$0, so microscopic stickiness appears macroscopically as reduced effective diffusivity (Aryasova et al., 15 Dec 2025). This suggests that sticky interfaces act not only as drift-generating membranes but also as decelerating media in homogenized limits.

4. Numerical approximation and simulation

Sticky boundary behavior is numerically delicate because the process changes effective dimension at the sticky set and standard time-stepping can miss the boundary residence mechanism. For multidimensional sticky diffusions, a continuous-time Markov chain approximation has been developed by approximating the generator either through finite-difference stencils in coordinate directions or by local-moment matching using the drift and the eigenvectors of the covariance matrix as jump directions (Meier et al., 2021). Both constructions yield first-order weak convergence: $\beta\in[-1,1]$1 and the method is described as free from the curse of dimensionality in simulation (Meier et al., 2021). The approach is illustrated on a two-dimensional sticky Brownian motion arising as a queuing limit and on a two-factor short-rate model with a sticky floor at zero (Meier et al., 2021).

In one dimension, a CTMC approximation on a grid $\beta\in[-1,1]$2 uses standard nearest-neighbor rates in the interior and special sticky-boundary rates at the lower endpoint $\beta\in[-1,1]$3. Two boundary schemes are given: $\beta\in[-1,1]$4 the second providing $\beta\in[-1,1]$5 accuracy away from payoff discontinuities (Meier et al., 2019). Matrix exponential methods then provide approximate Feynman–Kac operators and first-passage probabilities, and the same CTMC can be used for simulation. The paper emphasizes that the usual Euler–Maruyama discretization may completely fail in the very-sticky regime, whereas the CTMC captures boundary sticking robustly (Meier et al., 2019).

A particularly transparent approximation is the sticky random walk on $\beta\in[-1,1]$6. In the interior,

$\beta\in[-1,1]$7

while at the boundary

$\beta\in[-1,1]$8

Its mean holding time at $\beta\in[-1,1]$9 is $0$00, which is ballistic in $0$01 rather than diffusive, and this reproduces sticky residence in the continuum limit (Bou-Rabee et al., 2019). The resulting jump process converges weakly to sticky Brownian motion and, in the parameter regimes reported there, is two to five orders of magnitude faster than alternative methods (Bou-Rabee et al., 2019).

Recent work has also adapted Euler-type schemes to stickiness rather than abandoning them entirely. A modified Euler–Maruyama method for one-dimensional sticky diffusion separates a boundary layer $0$02 near zero. If the current state lies outside $0$03, one uses a reflected EM step; if it lies inside, one chooses between a reflected EM update and a jump to the sticky point $0$04 with a probability $0$05 derived by matching the one-step generator (Jiang et al., 25 Jun 2026). The scheme converges weakly with order $0$06 (Jiang et al., 25 Jun 2026). Complementary weak-approximation schemes of orders $0$07 and $0$08 have been constructed for general sticky boundary conditions via probabilistic representations of solutions to parabolic PDEs with second-order sticky boundary conditions (Sharma, 8 Aug 2025). A practical implication of this literature is that sticky diffusion is now numerically accessible both through generator-based CTMC methods and through boundary-corrected weak schemes, but not through naïve reflection-only discretizations.

5. High-frequency functionals, local time approximation, and parameter estimation

High-frequency observations of sticky diffusions support a distinct inference theory because the sticky parameter is tied to local time and occupation time rather than to drift or diffusion coefficients alone. For a sticky Itô diffusion with sticky point $0$09, a local-time proxy based on discrete observations at $0$10 is

$0$11

where $0$12 is bounded and integrable, vanishes near $0$13, and $0$14 with $0$15. Then

$0$16

locally uniformly in $0$17, in probability (Anagnostakis, 2022). Combined with a Riemann approximation of occupation time,

$0$18

this yields a consistent estimator of the stickiness parameter,

$0$19

conditionally on $0$20 (Anagnostakis, 2022).

The threshold-functional framework has since been generalized to sticky–oscillating–skew diffusions. For any normalizing sequence $0$21 with $0$22, the functional

$0$23

converges in ucp to

$0$24

where $0$25 is the speed measure of the associated sticky-skew Brownian motion (Anagnostakis et al., 2024). This extends the local-time approximation beyond the earlier regime $0$26 and supports consistent estimators not only of stickiness $0$27 but also of skewness $0$28 and side-specific volatilities $0$29 in sticky-threshold models (Anagnostakis et al., 2024).

A different inferential route counts threshold crossings. For sticky Brownian motion sampled on $0$30, three crossing statistics $0$31 exhibit three asymptotic scales: $0$32

$0$33

and the same trichotomy holds for corresponding bouncing counts (Anagnostakis et al., 2024). These limits lead to a consistent stickiness estimator

$0$34

on $0$35 (Anagnostakis et al., 2024). The coexistence of $0$36- and $0$37-scale statistics is specific to sticky thresholds and reflects the distinction between weak and closed-plane encounters with the sticky set.

6. Applications, boundary-value problems, and terminological extensions

A prominent applied instance is sticky geometric Brownian motion in mathematical finance. With stickiness level $0$38, parameter $0$39, risky asset $0$40, and riskless asset $0$41, the model is

$0$42

The model satisfies No Arbitrage and No Free Lunch with Vanishing Risk if and only if $0$43 (Anagnostakis, 2023). In that arbitrage-free case, the pricing function for a European payoff solves a mixed-boundary PDE: $0$44 together with the flux-matching condition

$0$45

(Anagnostakis, 2023). Continuous-time replication is still available through the delta $0$46, but the numerical evidence in that work indicates that larger stickiness increases discrete-time tracking error and that ignoring stickiness leads to a systematic PnL drift equal to the misprice (Anagnostakis, 2023).

Another direction links sticky diffusions to nonlocal boundary-value problems. For bounded smooth domains, an elastic-sticky Brownian motion obtained by slowing a reflecting Brownian motion through the clock

$0$47

has generator $0$48 with Wentzell–Robin boundary condition

$0$49

(D'Ovidio, 2022). Replacing the linear clock correction by a stable-subordinator term,

$0$50

yields a fractional sticky diffusion whose boundary condition becomes

$0$51

In that case the boundary holding times are Mittag–Leffler distributed and have infinite mean for $0$52, producing a macroscopic trap effect (D'Ovidio, 2022).

Sticky boundaries also arise in encounter-based reaction models. For sticky Brownian motion on the half-line, one considers the joint density $0$53 of particle position and boundary occupation time $0$54. Absorption is then defined as the first time $0$55 exceeds an independent random threshold $0$56, so the survival probability becomes

$0$57

where $0$58 (Bressloff, 2023). When $0$59 is exponential, the model recovers the classical constant-rate Robin mechanism; in the sticky case, however, the contact variable is true occupation time rather than merely boundary local time (Bressloff, 2023).

The term “sticky diffusion” is used more broadly in adjacent physical literature, and these usages are related but not identical. In lattice models of tracer motion among soft obstacles, “sticky” versus “slippery” refers to whether a bound tracer becomes immobile or remains mobile on obstacle sites; the long-time effective diffusion coefficient and transient anomalous exponent depend strongly on this distinction (Stefferson et al., 2017). In crowded polymer environments, attractive tracer–crowder interactions can generate non-monotonic diffusion as density varies, a phenomenon interpreted as competition between excluded volume and reversible adsorption (Nogueira et al., 2020). In a one-dimensional lattice gas combining exclusion dynamics with Ising-type nearest-neighbor attraction, “sticky diffusion” denotes a non-equilibrium transport model whose mean-field diffusion coefficient

$0$60

can change sign, signaling a transition from diffusive to blocked behavior (Hellier et al., 2018). These works suggest a broader terminological family centered on adhesion, dwell-time, or slowed release, even when the underlying mathematics is not the Wentzell/local-time formalism of continuous sticky diffusions.

Across these formulations, a common structural theme persists: stickiness is the insertion of a nontrivial time scale at a lower-dimensional set. Whether represented by an atom in the speed measure, a boundary term in a Dirichlet form, a local-time time change, a flux-matching condition in a PDE, or a discrete residence mechanism in a numerical or physical model, sticky diffusion differs from ordinary reflection precisely by converting instantaneous contact into measurable residence.