Threshold Diffusion Processes

• Threshold Diffusion Processes are stochastic models with piecewise-defined drift and diffusion coefficients, where crossing a threshold changes local dynamics and long-term behavior.
• They employ explicit scale–speed analysis, q-harmonic functions, and matching conditions to derive explicit results on hitting times, invariance, and escape probabilities.
• Applications span optimal stopping, first-passage problems, metastable state modeling, and network diffusion, illustrating how regime changes impact both statistical inference and dynamic evolution.

Threshold diffusion processes, in contemporary probability and stochastic-analysis literature, do not denote a single universally fixed model class. The term is used for several related constructions in which a level, interface, or domain boundary changes the local dynamics, the effective generator, or the relevant asymptotic regime. The central examples are one-dimensional diffusions with piecewise-defined drift and diffusion coefficients, multi-regime threshold Ornstein–Uhlenbeck models, singularly perturbed diffusions with trapping domains, and several adjacent notions in first-passage theory, optimal stopping, jump filtering, and diffusion-driven instability. What unifies these usages is that crossing a threshold modifies either the coefficients, the time-scale structure, or the inferential geometry of the model (Ji et al., 25 Aug 2025, Freidlin et al., 2015).

1. Definitions and principal model classes

A standard SDE-based meaning of a threshold diffusion is a one-dimensional diffusion whose coefficients are step functions. In the multi-regime formulation, thresholds

$-\infty<a_1<\cdots<a_n<\infty$

partition the state space, and the process is defined by

$$X_t = X_0 + \int_0^t b(X_s)\,ds + \int_0^t \sigma(X_s)\,dB_s,$$

with

$$\begin{cases} b(x) = \mu_0{\bf 1}_{(-\infty, a_1]}(x) + \displaystyle\sum_{i = 1}^{n-1}\mu_i {\bf 1}_{(a_i, a_{i+1}]}(x) + \mu_{n}{\bf 1}_{(a_n, \infty)}(x), \ \sigma(x) = \sigma_0{\bf 1}_{(-\infty, a_1]}(x) + \displaystyle\sum_{i = 1}^{n-1}\sigma_i {\bf 1}_{(a_i, a_{i+1}]}(x) + \sigma_{n}{\bf 1}_{(a_n, \infty)}(x), \end{cases}$$

where $\mu_i\in\mathbb R$ and $\sigma_i>0$. In that formulation, threshold diffusions are continuous-state, continuous-time analogues of threshold autoregressive or regime-switching models, with state-dependent rather than externally driven switching (Ji et al., 25 Aug 2025).

A particularly tractable single-threshold case is the drifted Oscillating Brownian motion. There the threshold is fixed at $0$, and both drift and diffusion coefficients are piecewise constant: $$\sigma(x)=\begin{cases} \sigma_+>0&\text{ if }x\geq 0,\ \sigma_->0&\text{ if }x<0, \end{cases} \qquad b(x)=\begin{cases} b_+\in \mathbb{R} &\text{ if }x\geq 0,\ b_-\in \mathbb{R} &\text{ if }x<0. \end{cases}$$ The process behaves as a Brownian motion with drift $b_+$ and volatility $\sigma_+$ on the positive side, and with drift $b_-$ and volatility $$X_t = X_0 + \int_0^t b(X_s)\,ds + \int_0^t \sigma(X_s)\,dB_s,$$0 on the negative side. The coefficients are discontinuous at the threshold in the literal sense that the right and left limits may differ (Lejay et al., 2018).

A further important class is the multi-regime threshold Ornstein–Uhlenbeck process,

$$X_t = X_0 + \int_0^t b(X_s)\,ds + \int_0^t \sigma(X_s)\,dB_s,$$1

with

$$X_t = X_0 + \int_0^t b(X_s)\,ds + \int_0^t \sigma(X_s)\,dB_s,$$2

Here the thresholds partition $$X_t = X_0 + \int_0^t b(X_s)\,ds + \int_0^t \sigma(X_s)\,dB_s,$$3 into $$X_t = X_0 + \int_0^t b(X_s)\,ds + \int_0^t \sigma(X_s)\,dB_s,$$4 regimes, and each regime carries its own affine mean-reverting drift $$X_t = X_0 + \int_0^t b(X_s)\,ds + \int_0^t \sigma(X_s)\,dB_s,$$5 (Han et al., 2024).

Usage in the literature	Canonical form	Representative source
Step-function threshold diffusion	Piecewise constant $$X_t = X_0 + \int_0^t b(X_s)\,ds + \int_0^t \sigma(X_s)\,dB_s,$$6, $$X_t = X_0 + \int_0^t b(X_s)\,ds + \int_0^t \sigma(X_s)\,dB_s,$$7 across thresholds	(Ji et al., 25 Aug 2025)
Drifted Oscillating Brownian motion	Two-regime discontinuity at $$X_t = X_0 + \int_0^t b(X_s)\,ds + \int_0^t \sigma(X_s)\,dB_s,$$8	(Lejay et al., 2018)
Threshold Ornstein–Uhlenbeck process	Piecewise affine drift on threshold intervals	(Han et al., 2024)

These SDE-based models form the core of the subject in the narrow sense. They are continuous-path diffusions with discontinuous coefficients, and they supply the main setting for explicit potential theory, recurrence classification, and asymptotic inference.

2. Generators, scale–speed analysis, and long-run regimes

For single-threshold models with piecewise constant coefficients, one-dimensional diffusion theory remains available in explicit form. For the drifted Oscillating Brownian motion, the infinitesimal generator can be written in divergence form as

$$X_t = X_0 + \int_0^t b(X_s)\,ds + \int_0^t \sigma(X_s)\,dB_s,$$9

with corresponding speed density and scale function

$$\begin{cases} b(x) = \mu_0{\bf 1}_{(-\infty, a_1]}(x) + \displaystyle\sum_{i = 1}^{n-1}\mu_i {\bf 1}_{(a_i, a_{i+1}]}(x) + \mu_{n}{\bf 1}_{(a_n, \infty)}(x), \ \sigma(x) = \sigma_0{\bf 1}_{(-\infty, a_1]}(x) + \displaystyle\sum_{i = 1}^{n-1}\sigma_i {\bf 1}_{(a_i, a_{i+1}]}(x) + \sigma_{n}{\bf 1}_{(a_n, \infty)}(x), \end{cases}$$0

These objects govern recurrence and transience. In that model the process is transient iff $$\begin{cases} b(x) = \mu_0{\bf 1}_{(-\infty, a_1]}(x) + \displaystyle\sum_{i = 1}^{n-1}\mu_i {\bf 1}_{(a_i, a_{i+1}]}(x) + \mu_{n}{\bf 1}_{(a_n, \infty)}(x), \ \sigma(x) = \sigma_0{\bf 1}_{(-\infty, a_1]}(x) + \displaystyle\sum_{i = 1}^{n-1}\sigma_i {\bf 1}_{(a_i, a_{i+1}]}(x) + \sigma_{n}{\bf 1}_{(a_n, \infty)}(x), \end{cases}$$1 or $$\begin{cases} b(x) = \mu_0{\bf 1}_{(-\infty, a_1]}(x) + \displaystyle\sum_{i = 1}^{n-1}\mu_i {\bf 1}_{(a_i, a_{i+1}]}(x) + \mu_{n}{\bf 1}_{(a_n, \infty)}(x), \ \sigma(x) = \sigma_0{\bf 1}_{(-\infty, a_1]}(x) + \displaystyle\sum_{i = 1}^{n-1}\sigma_i {\bf 1}_{(a_i, a_{i+1}]}(x) + \sigma_{n}{\bf 1}_{(a_n, \infty)}(x), \end{cases}$$2; it is ergodic iff $$\begin{cases} b(x) = \mu_0{\bf 1}_{(-\infty, a_1]}(x) + \displaystyle\sum_{i = 1}^{n-1}\mu_i {\bf 1}_{(a_i, a_{i+1}]}(x) + \mu_{n}{\bf 1}_{(a_n, \infty)}(x), \ \sigma(x) = \sigma_0{\bf 1}_{(-\infty, a_1]}(x) + \displaystyle\sum_{i = 1}^{n-1}\sigma_i {\bf 1}_{(a_i, a_{i+1}]}(x) + \sigma_{n}{\bf 1}_{(a_n, \infty)}(x), \end{cases}$$3 and $$\begin{cases} b(x) = \mu_0{\bf 1}_{(-\infty, a_1]}(x) + \displaystyle\sum_{i = 1}^{n-1}\mu_i {\bf 1}_{(a_i, a_{i+1}]}(x) + \mu_{n}{\bf 1}_{(a_n, \infty)}(x), \ \sigma(x) = \sigma_0{\bf 1}_{(-\infty, a_1]}(x) + \displaystyle\sum_{i = 1}^{n-1}\sigma_i {\bf 1}_{(a_i, a_{i+1}]}(x) + \sigma_{n}{\bf 1}_{(a_n, \infty)}(x), \end{cases}$$4; and the remaining recurrent cases are null recurrent (Lejay et al., 2018).

The same scale–speed structure appears in discretely observed threshold diffusions with piecewise constant coefficients. In the ergodic regime $$\begin{cases} b(x) = \mu_0{\bf 1}_{(-\infty, a_1]}(x) + \displaystyle\sum_{i = 1}^{n-1}\mu_i {\bf 1}_{(a_i, a_{i+1}]}(x) + \mu_{n}{\bf 1}_{(a_n, \infty)}(x), \ \sigma(x) = \sigma_0{\bf 1}_{(-\infty, a_1]}(x) + \displaystyle\sum_{i = 1}^{n-1}\sigma_i {\bf 1}_{(a_i, a_{i+1}]}(x) + \sigma_{n}{\bf 1}_{(a_n, \infty)}(x), \end{cases}$$5, the invariant distribution is the normalized speed measure, and the threshold acts as an attracting interface: the process is pushed upward below the threshold and downward above it. The resulting invariant density is piecewise exponential on the two half-lines, and stationary occupation probabilities on the positive and negative sides directly control regime-wise information accumulation in estimation problems (Mazzonetto et al., 2024).

The multi-regime step-function model admits a parallel but more global explicit analysis. On each interval the generator reduces to the constant-coefficient ODE

$$\begin{cases} b(x) = \mu_0{\bf 1}_{(-\infty, a_1]}(x) + \displaystyle\sum_{i = 1}^{n-1}\mu_i {\bf 1}_{(a_i, a_{i+1}]}(x) + \mu_{n}{\bf 1}_{(a_n, \infty)}(x), \ \sigma(x) = \sigma_0{\bf 1}_{(-\infty, a_1]}(x) + \displaystyle\sum_{i = 1}^{n-1}\sigma_i {\bf 1}_{(a_i, a_{i+1}]}(x) + \sigma_{n}{\bf 1}_{(a_n, \infty)}(x), \end{cases}$$6

and the paper constructs global increasing and decreasing $$\begin{cases} b(x) = \mu_0{\bf 1}_{(-\infty, a_1]}(x) + \displaystyle\sum_{i = 1}^{n-1}\mu_i {\bf 1}_{(a_i, a_{i+1}]}(x) + \mu_{n}{\bf 1}_{(a_n, \infty)}(x), \ \sigma(x) = \sigma_0{\bf 1}_{(-\infty, a_1]}(x) + \displaystyle\sum_{i = 1}^{n-1}\sigma_i {\bf 1}_{(a_i, a_{i+1}]}(x) + \sigma_{n}{\bf 1}_{(a_n, \infty)}(x), \end{cases}$$7-harmonic solutions $$\begin{cases} b(x) = \mu_0{\bf 1}_{(-\infty, a_1]}(x) + \displaystyle\sum_{i = 1}^{n-1}\mu_i {\bf 1}_{(a_i, a_{i+1}]}(x) + \mu_{n}{\bf 1}_{(a_n, \infty)}(x), \ \sigma(x) = \sigma_0{\bf 1}_{(-\infty, a_1]}(x) + \displaystyle\sum_{i = 1}^{n-1}\sigma_i {\bf 1}_{(a_i, a_{i+1}]}(x) + \sigma_{n}{\bf 1}_{(a_n, \infty)}(x), \end{cases}$$8 and $$\begin{cases} b(x) = \mu_0{\bf 1}_{(-\infty, a_1]}(x) + \displaystyle\sum_{i = 1}^{n-1}\mu_i {\bf 1}_{(a_i, a_{i+1}]}(x) + \mu_{n}{\bf 1}_{(a_n, \infty)}(x), \ \sigma(x) = \sigma_0{\bf 1}_{(-\infty, a_1]}(x) + \displaystyle\sum_{i = 1}^{n-1}\sigma_i {\bf 1}_{(a_i, a_{i+1}]}(x) + \sigma_{n}{\bf 1}_{(a_n, \infty)}(x), \end{cases}$$9 by stitching local exponentials across threshold points using $\mu_i\in\mathbb R$0 matching conditions. These functions yield explicit conditional Laplace transforms of hitting times: $\mu_i\in\mathbb R$1 as well as explicit potential measures, stationary distributions, and escape probabilities. Inward-pointing outer drifts,

$\mu_i\in\mathbb R$2

produce a stationary distribution; outward-pointing outer drifts,

$\mu_i\in\mathbb R$3

produce almost sure escape to $\mu_i\in\mathbb R$4 or $\mu_i\in\mathbb R$5 (Ji et al., 25 Aug 2025).

A general structural conclusion follows from these analyses. Threshold diffusions with discontinuous coefficients retain the classical one-dimensional toolkit—scale functions, speed measures, $\mu_i\in\mathbb R$6-harmonic functions, resolvents, and first-passage transforms—but every object becomes piecewise-defined and must satisfy interface matching conditions at the thresholds. The threshold therefore enters both locally, through coefficient jumps, and globally, through recurrence class, invariant law, and escape behavior.

3. Statistical inference and singular identification

Threshold parameters are statistically nonregular. For ergodic threshold diffusions with unknown switching values, the likelihood is not differentiable in the threshold parameter, the Fisher information is effectively infinite, and the threshold estimators converge at rate $\mu_i\in\mathbb R$7 rather than $\mu_i\in\mathbb R$8. In the asymptotics of large samples, the normalized likelihood ratios converge to exponential functionals of two-sided Wiener processes, and the maximum likelihood and Bayes estimators have different non-Gaussian limits. In Kutoyants’ formulation, threshold diffusion processes are direct continuous-time analogues of threshold autoregressive models, and the threshold-estimation problem is singular in exactly this sense (Kutoyants, 2010).

When the threshold is known and the drift coefficients are to be estimated, occupation times and local time become the dominant inferential objects. For the drifted Oscillating Brownian motion under continuous observation, the maximum likelihood estimator coincide with a quasi-likelihood estimator with constant diffusion term. Writing

$\mu_i\in\mathbb R$9

and

$\sigma_i>0$0

the estimator is

$\sigma_i>0$1

Using Tanaka’s formula,

$\sigma_i>0$2

The long-time behavior of the positive and negative occupation times rules the ones of the estimators, and this is why the asymptotic theory changes sharply across ergodic, null recurrent, and transient regimes (Lejay et al., 2018).

Discrete observation preserves the same structure, but the threshold crossing term must be approximated. In the ergodic two-regime model with piecewise constant coefficients, the generalized moment estimator

$\sigma_i>0$3

and the discretized maximum likelihood estimator

$\sigma_i>0$4

are linked by the exact identity

$\sigma_i>0$5

This yields asymptotic equivalence in the ergodic regime. In high frequency and infinite horizon, if

$\sigma_i>0$6

the asymptotic distribution of the discrete estimator matches that of the continuous-observation MLE (Mazzonetto et al., 2024).

For multi-regime threshold Ornstein–Uhlenbeck processes, least squares provides a unified alternative that extends beyond the two-regime case. Under continuous observation, the regime-wise least squares estimators of $\sigma_i>0$7 are strongly consistent and asymptotically normal with $\sigma_i>0$8 rate; under discrete observation with $\sigma_i>0$9, $0$0, and

$0$1

the same estimators are strongly consistent and asymptotically normal with $0$2 rate. The paper also proposes a modified quadratic variation estimator for the diffusion parameters and proves its consistency. A central practical implication is that the least squares method can be applied to the multi-regime cases $0$3, while the generalized moment method only deals with the two regime cases $0$4 (Han et al., 2024).

Taken together, these results show that threshold diffusion inference has two distinct asymptotic geometries. Unknown threshold locations generate singular $0$5-rate estimation problems, whereas known-threshold coefficient estimation is regular once regime occupation times are properly accounted for. Local time at the threshold is the mechanism linking the two.

4. Thresholds as stopping and first-passage boundaries

A second major meaning of threshold in diffusion theory concerns stopping and passage rather than coefficient discontinuity. In one-dimensional diffusion optimal stopping, a threshold strategy is a stopping time of the form “stop the first time the process exceeds a level.” For a regular diffusion on an interval $0$6, with generator

$0$7

and increasing fundamental solution $0$8 of

$0$9

the expected discounted payoff under an upper-threshold rule has the form

$$\sigma(x)=\begin{cases} \sigma_+>0&\text{ if }x\geq 0,\ \sigma_->0&\text{ if }x<0, \end{cases} \qquad b(x)=\begin{cases} b_+\in \mathbb{R} &\text{ if }x\geq 0,\ b_-\in \mathbb{R} &\text{ if }x<0. \end{cases}$$0

The threshold stopping time $$\sigma(x)=\begin{cases} \sigma_+>0&\text{ if }x\geq 0,\ \sigma_->0&\text{ if }x<0, \end{cases} \qquad b(x)=\begin{cases} b_+\in \mathbb{R} &\text{ if }x\geq 0,\ b_-\in \mathbb{R} &\text{ if }x<0. \end{cases}$$1 is optimal among threshold rules iff $$\sigma(x)=\begin{cases} \sigma_+>0&\text{ if }x\geq 0,\ \sigma_->0&\text{ if }x<0, \end{cases} \qquad b(x)=\begin{cases} b_+\in \mathbb{R} &\text{ if }x\geq 0,\ b_-\in \mathbb{R} &\text{ if }x<0. \end{cases}$$2 whenever $$\sigma(x)=\begin{cases} \sigma_+>0&\text{ if }x\geq 0,\ \sigma_->0&\text{ if }x<0, \end{cases} \qquad b(x)=\begin{cases} b_+\in \mathbb{R} &\text{ if }x\geq 0,\ b_-\in \mathbb{R} &\text{ if }x<0. \end{cases}$$3 and $$\sigma(x)=\begin{cases} \sigma_+>0&\text{ if }x\geq 0,\ \sigma_->0&\text{ if }x<0, \end{cases} \qquad b(x)=\begin{cases} b_+\in \mathbb{R} &\text{ if }x\geq 0,\ b_-\in \mathbb{R} &\text{ if }x<0. \end{cases}$$4 does not increase for $$\sigma(x)=\begin{cases} \sigma_+>0&\text{ if }x\geq 0,\ \sigma_->0&\text{ if }x<0, \end{cases} \qquad b(x)=\begin{cases} b_+\in \mathbb{R} &\text{ if }x\geq 0,\ b_-\in \mathbb{R} &\text{ if }x<0. \end{cases}$$5. In the full stopping problem, this must be supplemented by the condition $$\sigma(x)=\begin{cases} \sigma_+>0&\text{ if }x\geq 0,\ \sigma_->0&\text{ if }x<0, \end{cases} \qquad b(x)=\begin{cases} b_+\in \mathbb{R} &\text{ if }x\geq 0,\ b_-\in \mathbb{R} &\text{ if }x<0. \end{cases}$$6 on the stopping side and a one-sided derivative inequality at $$\sigma(x)=\begin{cases} \sigma_+>0&\text{ if }x\geq 0,\ \sigma_->0&\text{ if }x<0, \end{cases} \qquad b(x)=\begin{cases} b_+\in \mathbb{R} &\text{ if }x\geq 0,\ b_-\in \mathbb{R} &\text{ if }x<0. \end{cases}$$7 (Arkin et al., 2013).

The variational formulation sharpens the same point. In that approach, smooth pasting is not a verification principle by itself; it is the first-order stationarity condition of the one-parameter optimization problem for $$\sigma(x)=\begin{cases} \sigma_+>0&\text{ if }x\geq 0,\ \sigma_->0&\text{ if }x<0, \end{cases} \qquad b(x)=\begin{cases} b_+\in \mathbb{R} &\text{ if }x\geq 0,\ b_-\in \mathbb{R} &\text{ if }x<0. \end{cases}$$8 or $$\sigma(x)=\begin{cases} \sigma_+>0&\text{ if }x\geq 0,\ \sigma_->0&\text{ if }x<0, \end{cases} \qquad b(x)=\begin{cases} b_+\in \mathbb{R} &\text{ if }x\geq 0,\ b_-\in \mathbb{R} &\text{ if }x<0. \end{cases}$$9. A solution of the free-boundary problem can therefore be spurious: it may correspond to a stationary point of the threshold objective rather than a true maximizer. The paper gives explicit second-order criteria for discarding such free-boundary candidates (Arkin et al., 2015).

Thresholds also define first-passage and first-exit times in classical drift–diffusion models. In the one-dimensional drift–diffusion equation

$b_+$0

decision time in the double-threshold model is the first exit time from $b_+$1,

$b_+$2

and the paper derives explicit formulas for the first three moments of the decision time distribution, together with moments conditioned on hitting the upper or lower threshold (Srivastava et al., 2016).

A related first-passage literature studies explicitly time-dependent barriers. For two stochastic versions of a general growth curve, one with multiplicative noise and one with additive noise, the paper identifies classes of time-dependent thresholds for which the first passage time from a barrier and the first exit time from a region delimited by two barriers admit closed-form densities. The key reduction is to Gauss–Markov processes, and in the multiplicative-noise case to a Wiener process after a logarithmic transformation (Albano et al., 2024).

These stopping and passage problems are not threshold diffusions in the narrow coefficient-switching sense. They are, however, part of the same conceptual family: the threshold is the geometric locus at which either the dynamics terminate, the payoff switches, or the observable event is recorded.

5. Geometric and metastable threshold mechanisms

A different threshold mechanism arises when entering a domain activates a very large inward drift. On the $b_+$3-dimensional torus $b_+$4, let

$b_+$5

be pairwise disjoint domains, define

$b_+$6

and consider

$b_+$7

where $b_+$8 vanishes on $b_+$9 and points inward on each $\sigma_+$0. Outside the domains the process is Brownian motion; inside them the drift is of order $\sigma_+$1. As $\sigma_+$2, each $\sigma_+$3 becomes trapping, and the time to exit the domains is exponentially large. The threshold is not a piecewise switch in the usual sense; it is the boundary of a metastable region, and crossing it changes the qualitative time-scale structure of the process (Freidlin et al., 2015).

The ordinary-time limit is obtained by suppressing the exponentially long trap sojourns and passing to the trace process on the quotient space $\sigma_+$4 obtained by collapsing each $\sigma_+$5 to a point $\sigma_+$6. The limiting generator acts as $\sigma_+$7 in $\sigma_+$8, but its domain carries the nonstandard boundary condition

$\sigma_+$9

with

$b_-$0

Thus the effective boundary behavior is neither reflecting nor absorbing. On exponential time scales, the same model yields metastable distributions between traps, determined by harmonic boundary-value problems with integral flux conditions (Freidlin et al., 2015).

Graph-limit constructions provide another geometric threshold mechanism. Diffusions in tubular domains shrinking to curves, spiders, and general graphs converge to graph diffusions characterized by second-order generators on the limiting graph. Along each edge the limit is a one-dimensional diffusion, but at each vertex the process satisfies a weighted Kirchhoff condition. In the $b_-$1-spider case the domain of the generator contains the gluing rule

$b_-$2

and the edge-selection weights are

$b_-$3

In the reflecting case the weights simplify to

$b_-$4

A plausible implication is that the junction acts as a microscopic threshold: after collapse of the transverse geometry, local tube widths and confining potentials become probabilistic branch-selection rules at the vertex (Albeverio et al., 2012).

Both the trap model and the thin-tube limit show that a threshold need not be a scalar level in one dimension. It may be a hypersurface or junction at which the effective Markov structure changes from ordinary diffusion to metastable recycling, or from Euclidean motion to graph transmission with weighted Kirchhoff coupling.

6. Broader diffusion-related meanings of threshold

The literature also uses threshold language in several broader diffusion-related settings. In a competitive contact process on a bipartite square lattice, diffusion is a homogenizing mechanism that destroys sublattice ordering above a threshold rate. For low diffusion, the model exhibits an active asymmetric phase with $b_-$5; above a threshold diffusion the sublattice ordering is suppressed and only the usual active symmetric–inactive transition remains. In mean-field form, diffusion enters the symmetry-breaking equation through the term $b_-$6, so diffusion acts directly as a damping term on sublattice imbalance (Oliveira et al., 2017).

In random reaction–diffusion systems, the diffusive threshold is the minimum disparity in diffusivities required for deterministic Turing instability. For $b_-$7 diffusing species this threshold is typically substantial except under kinetic fine tuning, whereas for $b_-$8 the paper finds that the diffusive threshold becomes more likely to be smaller and physical as $b_-$9 increases, at least in the numerically tractable cases $$X_t = X_0 + \int_0^t b(X_s)\,ds + \int_0^t \sigma(X_s)\,dB_s,$$00 (Haas et al., 2020).

In high-frequency statistics for jump-diffusions, threshold is often an inference device rather than a dynamical feature. Local polynomial threshold estimators of the volatility function use the indicator

$$X_t = X_0 + \int_0^t b(X_s)\,ds + \int_0^t \sigma(X_s)\,dB_s,$$01

so that increments larger than the threshold are filtered out as jump-contaminated. The same idea appears in high-frequency structural equation models for jump-diffusions, where increments are retained when

$$X_t = X_0 + \int_0^t b(X_s)\,ds + \int_0^t \sigma(X_s)\,dB_s,$$02

and the thresholded increments define a Gaussian quasi-likelihood. In both cases, thresholding is purely a jump-removal mechanism; it is not a threshold in the state dynamics themselves (Song et al., 2017, Kusano et al., 19 May 2025).

A network-theoretic extension replaces a scalar epidemic threshold by a multidimensional threshold set. For diffusion over interdependent networks with distinct node sets and different within-layer and inter-layer transmission rates, the threshold is a set of minimal transmission-rate vectors. The onset criterion is spectral: a multilayer epidemic occurs when the largest eigenvalue of the Jacobian matrix of the occupied diffusion network exceeds one (Salehi et al., 2014).

These broader usages clarify an important terminological point. In current research, threshold diffusion may refer to at least three non-equivalent ideas: a diffusion whose coefficients switch across a state threshold; a diffusion-related system controlled by a critical rate or instability boundary; or a statistical procedure that thresholds increments to isolate the diffusion part of jump-contaminated data. The precise meaning is therefore model-dependent, but in every case the threshold marks the locus where the qualitative description changes.

Across these literatures, threshold diffusion processes are best understood as a family of models and methods rather than a single canonical object. The most classical member of that family is the one-dimensional diffusion with discontinuous coefficients at fixed levels. Around it lie several mathematically adjacent structures: threshold Ornstein–Uhlenbeck dynamics, singular threshold estimation, first-threshold passage problems, trap-induced metastability, graph-junction transmission laws, and threshold filters for jump-robust inference. The subject is unified not by one formula, but by a recurring analytic theme: thresholds create interfaces at which local diffusion behavior, global asymptotics, or statistical identifiability change abruptly.