Modified Tamed Scheme Techniques for SDEs

• Modified tamed scheme is a class of numerical discretizations that refines classical taming techniques to stabilize SDEs with superlinear coefficients and low regularity.
• They employ adaptive time-stepping, Milstein derivative corrections, and selective modifications like far-field cut-offs to preserve order accuracy.
• These methods improve strong convergence rates and extend to structured problems including SDAEs, McKean–Vlasov equations, and Langevin dynamics.

Searching arXiv for papers on modified/tamed schemes to ground the article in current literature. arxiv_search(query="modified tamed scheme stochastic differential equations", max_results=10) to=arxiv_search 荣富? {"query":"modified tamed scheme stochastic differential equations", "max_results": 10} Modified tamed scheme denotes a class of numerical discretizations in which the basic taming device used for nonglobally Lipschitz stochastic dynamics is altered so that explicit or semi-implicit methods remain stable under superlinear growth, low regularity, singular interactions, algebraic constraints, or long-time sampling requirements. In the cited literature, the modification may take the form of adaptive time stepping, a tamed Milstein correction, a smooth cut-off that activates taming only in the far field, a truncation near a singularity set, or a problem-specific implicit-explicit split. These constructions are used for strong approximation of SDEs, SDAEs, McKean–Vlasov equations, and related Langevin samplers, with convergence orders ranging from the Euler-type rate $1/2$ to Milstein-type order $1$, and with additional results on infinite-horizon error bounds, geometric ergodicity, and relative-entropy control (Vu et al., 2024, Ju et al., 13 Jul 2025, Bao et al., 2024).

1. Origins in taming and explicit stabilization

The conceptual starting point is the failure of classical explicit discretizations for SDEs with superlinearly growing coefficients. For such equations, the explicit Euler method can fail to converge strongly, and standard explicit Milstein analysis typically requires substantially more regularity than is available in many nonglobally Lipschitz models. Taming was introduced to preserve explicit implementability while suppressing increments that would otherwise become too large. In the tamed Euler setting, this restores strong order $1/2$ in nonstandard problems such as stochastic differential equations with piecewise continuous arguments, while Milstein-type taming can recover higher-order strong accuracy in commutative-noise settings (Song et al., 2015, Wang et al., 2011).

Within this development, the adjective “modified” does not designate a single canonical algorithm. Rather, it refers to alterations of the standard taming mechanism that are tailored to a specific obstruction: low differentiability of coefficients, singular or constrained geometry, superlinear drift combined with long-time ergodicity, or loss of weak order under uniform denominator-based taming. A general abstract version appears in the framework of modified Euler recursions

$$X_{(n+1)\delta}=T^{(\delta)}(X_{n\delta})+b^{(\delta)}\!\bigl(T^{(\delta)}(X_{n\delta})\bigr)\delta+\sqrt{\delta}\,\xi_{n+1},$$

where $T^{(\delta)}$ is contractive and $b^{(\delta)}\to b$ as $\delta\to 0$. This framework explicitly encompasses the tamed Euler scheme and the truncated Euler scheme (Bao et al., 2024).

2. Principal forms of modification

The literature uses several distinct mechanisms to modify a tamed scheme. They differ not only in implementation but in the analytical role played by the modification.

Variant	Defining modification	Representative paper
Adaptive tamed Milstein	Adaptive step size and tamed Milstein derivative term $q_\Delta$	(Vu et al., 2024)
Far-field cut-off taming	$b^h=b/(1+\psi(\gamma h^\alpha |b|))$, with no taming for small $|b|$	(Ju et al., 13 Jul 2025)
Semi-implicit taming	Linear drift implicit, nonlinear drift explicitly tamed	(Tsafack et al., 10 Sep 2025)
Singular-set cut-off	Drift turned off when $1$0 is too small	(Johnston et al., 2024)
Contractive modified Euler	Contractive state map $1$1 and modified drift $1$2	(Bao et al., 2024)

These variants share the same strategic objective: preserve explicitness or near-explicitness while controlling excursions that invalidate classical convergence theory. Their differences are structural. Adaptive Milstein schemes modify both local step length and Milstein correction. Semi-implicit schemes use taming only on the nonlinear component, leaving linear stiff structure implicit. Singular-drift schemes replace denominator taming by a geometric cut-off tied to distance from a singular set. Far-field schemes use a smooth cut-off so that the discretization coincides exactly with the untamed method on the region where the drift is moderate.

This diversity has an important methodological consequence. A modified tamed scheme is best understood not as a single formula but as a design principle: alter the taming only where instability or lack of regularity actually enters the proof.

3. Adaptive Milstein under low regularity

A particularly sharp realization appears in the tamed-adaptive Milstein scheme for the one-dimensional SDE

$1$3

The novelty is that the coefficients are assumed to satisfy only

$1$4

meaning that the first derivatives are locally Hölder continuous of order $1$5 with polynomial growth, rather than globally Lipschitz. The scheme is defined by

$1$6

and

$1$7

with tamed Milstein coefficient

$1$8

The adaptive step-size function is

$1$9

Relative to classical Milstein, the method therefore changes both the step size and the derivative correction. Relative to standard tamed Milstein, it is explicitly tamed-adaptive rather than merely coefficient-tamed (Vu et al., 2024).

The assumptions are organized as $1/2$0–$1/2$1. Besides $1/2$2-regularity, they impose a coercivity-type moment condition

$1/2$3

and a one-sided stability condition

$1/2$4

The local low-regularity structure enters through the remainder estimate

$1/2$5

Under these assumptions, if

$1/2$6

then for every $1/2$7,

$1/2$8

Hence the $1/2$9-strong convergence rate is

$$X_{(n+1)\delta}=T^{(\delta)}(X_{n\delta})+b^{(\delta)}\!\bigl(T^{(\delta)}(X_{n\delta})\bigr)\delta+\sqrt{\delta}\,\xi_{n+1},$$0

If $$X_{(n+1)\delta}=T^{(\delta)}(X_{n\delta})+b^{(\delta)}\!\bigl(T^{(\delta)}(X_{n\delta})\bigr)\delta+\sqrt{\delta}\,\xi_{n+1},$$1 and $$X_{(n+1)\delta}=T^{(\delta)}(X_{n\delta})+b^{(\delta)}\!\bigl(T^{(\delta)}(X_{n\delta})\bigr)\delta+\sqrt{\delta}\,\xi_{n+1},$$2, the same estimate holds uniformly on the infinite horizon,

$$X_{(n+1)\delta}=T^{(\delta)}(X_{n\delta})+b^{(\delta)}\!\bigl(T^{(\delta)}(X_{n\delta})\bigr)\delta+\sqrt{\delta}\,\xi_{n+1},$$3

with $$X_{(n+1)\delta}=T^{(\delta)}(X_{n\delta})+b^{(\delta)}\!\bigl(T^{(\delta)}(X_{n\delta})\bigr)\delta+\sqrt{\delta}\,\xi_{n+1},$$4 independent of $$X_{(n+1)\delta}=T^{(\delta)}(X_{n\delta})+b^{(\delta)}\!\bigl(T^{(\delta)}(X_{n\delta})\bigr)\delta+\sqrt{\delta}\,\xi_{n+1},$$5. The same rate is therefore valid on both finite intervals and $$X_{(n+1)\delta}=T^{(\delta)}(X_{n\delta})+b^{(\delta)}\!\bigl(T^{(\delta)}(X_{n\delta})\bigr)\delta+\sqrt{\delta}\,\xi_{n+1},$$6 (Vu et al., 2024).

The proof combines moment bounds, adaptive-step control showing $$X_{(n+1)\delta}=T^{(\delta)}(X_{n\delta})+b^{(\delta)}\!\bigl(T^{(\delta)}(X_{n\delta})\bigr)\delta+\sqrt{\delta}\,\xi_{n+1},$$7 almost surely, increment estimates, a Yamada–Watanabe smoothing argument, and a Gronwall-type estimate for the error process. A central estimate is

$$X_{(n+1)\delta}=T^{(\delta)}(X_{n\delta})+b^{(\delta)}\!\bigl(T^{(\delta)}(X_{n\delta})\bigr)\delta+\sqrt{\delta}\,\xi_{n+1},$$8

which stabilizes the Milstein correction under only local Hölder continuity of derivatives. The resulting rate interpolates directly with the smoothness index $$X_{(n+1)\delta}=T^{(\delta)}(X_{n\delta})+b^{(\delta)}\!\bigl(T^{(\delta)}(X_{n\delta})\bigr)\delta+\sqrt{\delta}\,\xi_{n+1},$$9: smoother derivatives improve the strong order, but global Lipschitz differentiability is no longer required.

4. Far-field taming and preservation of Euler order

A different modification is designed to remove a specific weakness of standard denominator taming. In the usual tamed Euler construction,

$T^{(\delta)}$0

the drift perturbation is present everywhere, including regions where $T^{(\delta)}$1 is moderate. The modified tamed scheme replaces this by

$T^{(\delta)}$2

where $T^{(\delta)}$3 is smooth and satisfies

$T^{(\delta)}$4

with smooth increasing interpolation for $T^{(\delta)}$5. Consequently, if $T^{(\delta)}$6, then

$T^{(\delta)}$7

so no taming occurs in the bulk region. In the far field one recovers a standard tamed-type denominator, and the bound

$T^{(\delta)}$8

still provides stability (Ju et al., 13 Jul 2025).

For the Euler discretization,

$T^{(\delta)}$9

this selective taming changes the error mechanism. The paper proves the high-order estimate

$b^{(\delta)}\to b$0

for any $b^{(\delta)}\to b$1. The analytical reason is that taming is activated only on the event $b^{(\delta)}\to b$2, whose probability is made arbitrarily small by moment bounds. This removes the built-in low-order perturbation that standard taming introduces everywhere (Ju et al., 13 Jul 2025).

The resulting convergence theory recovers the underlying Euler orders. Under the strong assumptions,

$b^{(\delta)}\to b$3

so the modified tamed Euler scheme has strong order $b^{(\delta)}\to b$4. Under the weak assumptions, for any $b^{(\delta)}\to b$5,

$b^{(\delta)}\to b$6

so the weak order is $b^{(\delta)}\to b$7. The paper explicitly contrasts this with standard tamed schemes, which are described as introducing an $b^{(\delta)}\to b$8-type perturbation that can lower weak order (Ju et al., 13 Jul 2025).

The same construction is extended to random-batch approximations and to stochastic gradient Langevin dynamics. For the Langevin application, the paper proves the uniform-in-time relative-entropy bound

$b^{(\delta)}\to b$9

for any $\delta\to 0$0 with $\delta\to 0$1. This yields a uniform-in-time near-sharp error estimate under relative entropy for super-linear sampling dynamics (Ju et al., 13 Jul 2025).

5. Structured, constrained, and singular variants

Problem-specific modifications become especially pronounced once the underlying stochastic model is no longer a standard SDE. For index-1 SDAEs with a singular matrix $\delta\to 0$2, the modified construction is semi-implicit rather than fully explicit. The update

$\delta\to 0$3

treats the linear drift part implicitly, the nonlinear drift part explicitly but tamed, and the diffusion explicitly. A dual tamed scheme in reduced $\delta\to 0$4-variables is then used for analysis, and the continuous-time interpolation satisfies

$\delta\to 0$5

The modification is therefore dictated by singular structure and algebraic constraints rather than by low regularity alone (Tsafack et al., 10 Sep 2025).

For additive-noise SDEs with drift that is not locally integrable near a singularity set $\delta\to 0$6, denominator taming is replaced by a geometric cut-off. Writing $\delta\to 0$7, the modified drift is

$\delta\to 0$8

Thus the drift is turned off whenever the scheme approaches the singular set too closely. The resulting explicit Euler-type approximation converges strongly in $\delta\to 0$9 with rate $q_\Delta$0,

$q_\Delta$1

and the classical Euler scheme is shown not even to possess finite first moments in this regime (Johnston et al., 2024).

Further extensions show that the same design principle survives in more structured stochastic systems. For McKean–Vlasov equations with common noise, explicit tamed Euler and tamed Milstein schemes are built for the interacting particle system, with denominator taming on both drift and diffusion and with measure-derivative Milstein corrections; the strong orders are $q_\Delta$2 and $q_\Delta$3, respectively. For regime-switching SDEs, an explicit tamed Milstein-type scheme with an additional switching correction achieves strong $q_\Delta$4 order $q_\Delta$5. For neutral stochastic differential delay equations, tamed EM and tamed $q_\Delta$6-EM schemes provide strong convergence and, in some settings, exponential mean-square stability and almost sure exponential stability (Kumar et al., 2020, Kumar et al., 2019, Ji et al., 2016, Tan et al., 2016, Ji et al., 2019).

6. Ergodicity, sampling, and recurrent misconceptions

Modified taming is not limited to finite-time pathwise approximation. In the study of invariant measures and long-time dynamics, modified Euler recursions have been shown to be geometrically ergodic under a mixed probability distance and under weighted total variation distance. For the tamed Euler specialization,

$q_\Delta$7

one also obtains strict $q_\Delta$8-Wasserstein contractivity,

$q_\Delta$9

and the invariant measures satisfy

$b^h=b/(1+\psi(\gamma h^\alpha |b|))$0

These results use refined basic coupling for the general modified framework and coupling by reflection for the tamed Euler scheme (Bao et al., 2024).

In Langevin sampling, the modified Tamed Unadjusted Langevin Algorithm replaces the drift by

$b^h=b/(1+\psi(\gamma h^\alpha |b|))$1

and the chain

$b^h=b/(1+\psi(\gamma h^\alpha |b|))$2

admits non-asymptotic convergence bounds in Wasserstein distance under non-convexity and super-linear growth. The discretization bias is $b^h=b/(1+\psi(\gamma h^\alpha |b|))$3 in $b^h=b/(1+\psi(\gamma h^\alpha |b|))$4 and $b^h=b/(1+\psi(\gamma h^\alpha |b|))$5 in $b^h=b/(1+\psi(\gamma h^\alpha |b|))$6, with the analysis relying on dissipativity, Lyapunov drift estimates, and contractivity of a weighted semimetric rather than global strong convexity (Neufeld et al., 2022).

Two recurrent misconceptions are corrected by this body of work. First, taming does not mean only the classical denominator $b^h=b/(1+\psi(\gamma h^\alpha |b|))$7. In the cited literature it may mean a tamed Milstein derivative correction, a smooth far-field cut-off, an implicit-explicit split, a contractive state map, or a singular-set truncation. Second, taming does not inherently force a loss of the underlying discretization order. The adaptive Milstein construction achieves rate $b^h=b/(1+\psi(\gamma h^\alpha |b|))$8 under only $b^h=b/(1+\psi(\gamma h^\alpha |b|))$9 regularity, the selective far-field modification retains weak order $|b|$0 for Euler, and several Milstein-type tamed methods recover order $|b|$1 in commutative, mean-field, or switching settings (Vu et al., 2024, Ju et al., 13 Jul 2025, Wang et al., 2011, Kumar et al., 2020, Kumar et al., 2019).

Taken together, these results define the modified tamed scheme as a flexible analytical and algorithmic template rather than a single method. Its essential feature is selective stabilization: the discretization is altered exactly where superlinearity, irregularity, constraint structure, or long-time sampling analysis requires intervention, and left unchanged where the original explicit method already has the correct local behavior.