Tamed Euler Schemes

Updated 9 March 2026

Tamed Euler schemes are explicit numerical methods that stabilize SDEs with superlinear or singular coefficients by modifying drift and diffusion terms.
They use damping denominators or truncations to preserve moment stability and achieve strong convergence, typically at an order of 1/2.
These schemes extend to high-dimensional, delayed, and jump-driven systems, offering robust performance where classical Euler-Maruyama fails.

A tamed Euler scheme is an explicit numerical discretization method for stochastic or deterministic differential equations with coefficients that exhibit superlinear growth or singularities, for which the classical Euler–Maruyama method is unstable due to divergence of moments or blow-up. Taming modifies the drift and/or diffusion (and possibly jump) coefficients by introducing nonlinearity-damping denominators or truncations, thereby preserving stability and enabling rigorous convergence analysis for otherwise intractable problems. Contemporary research has generalized taming to SDEs, SPDEs, McKean-Vlasov systems (with and without common or Lévy noise), and equations with various forms of singularity, delay, regime switching, or time- and measure-dependence.

1. Mathematical Formulation and Core Principles

Let $X(t)$ solve an SDE,

$dX(t) = b(t, X(t))\,dt + \sigma(t, X(t))\,dW_t,\qquad X(0) = X_0,$

where $b$ may grow faster than linearly in $x$ , and $\sigma$ may be globally Lipschitz or similarly superlinear.

The foundational tamed Euler method replaces the drift with a bounded form,

$b_h(t, x) = \frac{b(t, x)}{1 + h^\alpha |b(t, x)|},$

where $h$ is the timestep, and $\alpha\in(0,1/2]$ . The explicit update at grid points $t_k = kh$ is

$X_{k+1} = X_k + h\,b_h(t_k, X_k) + \sigma(t_k, X_k)\,\Delta W_k,$

where $\Delta W_k = W_{t_{k+1}} - W_{t_k}$ (Sabanis, 2013).

For higher accuracy or improved control, modified tamed schemes may:

Use cutoff functions or state-dependent truncations (Ju et al., 13 Jul 2025).
Tame only when $|b|$ exceeds a threshold, using smooth cutoffs.
Apply taming to the measure argument or to polynomial functions of the state (Soni et al., 18 Oct 2025).

For stochastic delay, jump-driven, or neutral equations, analogous tamed versions replace each potentially unbounded coefficient with a tamed or truncated form (Ji et al., 2016, Tambue et al., 2015).

2. Theoretical Guarantees: Convergence, Stability, and Ergodicity

Strong and Weak Convergence

If $b$ is locally one-sided Lipschitz with polynomial growth, and $\alpha=1/2$ , the tamed Euler method achieves strong convergence order $1/2$ in $L^p$ for all $p>0$ (i.e., global error $O(h^{1/2})$ ) (Sabanis, 2013, Hu et al., 2022, Tambue et al., 2015): $\sup_{0\leq t\leq T} \mathbb{E}|X(t) - X_h(t)|^p \leq C h^{p/2}.$ Weak convergence rates are also established under additional smoothness (Ju et al., 13 Jul 2025).

Numerous generalizations attain similar rates:

For modified tamed schemes with smooth cutoffs, moment stability and $O(h^{1/2})$ error are proven, with bias from taming made arbitrarily small without loss of convergence order (Ju et al., 13 Jul 2025).
In superlinear measure-dependent setups (e.g., McKean-Vlasov), strong propagation of chaos and convergence with particle error $O(N^{-1/2})$ (for $N$ particles) and time-discretization error $O(h^{1/2})$ are achieved (Soni et al., 18 Oct 2025, Kumar et al., 2020).
Schemes for SDEs with Lévy noise, singular interaction kernels, Carathéodory drift, or non-Locally integrable drift admit $L^p$ -error rates $1/2$ under monotonicity and suitable taming (Biswas et al., 21 Oct 2025, Johnston et al., 2024).

Stability, Invariant Measures, and Ergodicity

Tamed schemes yield:

Uniform moment bounds for all moments of the numerical solution (Sabanis, 2013).
Exponential (geometric) ergodicity of the Markov chain in Wasserstein or weighted total variation distance, with explicit rates depending on the taming exponent (Bao et al., 2024, Liu et al., 2024).
Quantitative Wasserstein error between the true invariant measure and the scheme’s invariant measure; specifically, for tamed Euler with drift $b$ and timestep $\delta$ ,

$W_1(\pi, \pi^\delta) \leq C \delta^\theta,$

where $\theta$ is the taming exponent (Bao et al., 2024).

Stability—mean-square exponential decay—for both SDEs and SDDEs, including with jumps (Tambue et al., 2015, Tambue et al., 2015).

3. Extensions: Nonstandard Equations and Coefficient Structures

SDEs with Discontinuities and Singularities

Piecewise continuous or discontinuous drift and diffusions (with nonzero values at discontinuities) admit strong $L^p$ -$1/2$-rate tamed discretizations (Hu et al., 2022, Song et al., 2015).
Singular repulsive particle systems and mean-field SDEs with boundary blowup can be handled using domain-dependent taming, Lyapunov-dominated drift controls, or state-dependent truncations (Johnston et al., 23 Jan 2026, Johnston et al., 2024).

SPDEs and Stochastic Evolution Equations

In the infinite-dimensional Gelfand triple setting, tamed Euler–Galerkin methods yield stability and convergence for SPDEs with super-linear drift (no explicit strong rate, but convergence by compactness, moment control, and monotonicity) (Gyöngy et al., 2014).
Polygonal Euler schemes with monotonicity-preserving truncation yield explicit methods even in Hilbert space, with strong convergence rates depending on drift polynomiality (Johnston et al., 2021).

Equations with Delays, Jumps, or Switching

Tamed and semi-tamed schemes are extended to stochastic jump SDEs, delay equations, and Markovian switching, giving strong $1/2$-order convergence and exponential stability (Ji et al., 2016, Tambue et al., 2015, Neelima et al., 2020).

McKean-Vlasov Equations (with or without Lévy noise/Common noise)

Fully explicit tamed Euler schemes are developed for interacting particle systems whose coefficients depend superlinearly on both particle state and empirical measure. Quantitative propagation of chaos and ergodicity for the empirical law are established, including under non-globally Lipschitz coefficients and for common/Lévy noise (Soni et al., 18 Oct 2025, Neelima et al., 2020, Kumar et al., 2020).

4. Algorithmic Aspects and Practical Implementation

Method	Tamed Formulation	Key Property
Classical tamed Euler	$f_h(x) = f(x)/(1 + h^\alpha \|f(x)\|)$	Explicit, order $1/2$
Modified tamed Euler	$b^h = b/(1 + \psi(\gamma h^\alpha \|b\|))$	Improved bias, $O(h^{1/2})$
Polygonal truncation	$T_{R, r, x_0}[f]$ as piecewise combination	Monotonicity preserved
Tamed for jump SDEs	$f_h(x) = f(x)/(1 + \Delta t^\alpha \|f(x)\|)$	Handles jumps, $1/2$ rate
Randomized tamed Euler	Drift tamed/tamed randomization	Carathéodory drift
Lyapunov-tamed Euler	$b_h(x)=b(x)\mathbf{1}_{V(x)\le h^{-1/2}}$	Singular drift, $1/2$/$1$
Tamed EM for Neutral SDDE	$b_h(x, y) = b(x, y)/(1 + h\,\|b(x, y)\|)$	Neutral delays, $1/2$ rate

All these schemes can be implemented fully explicitly, per-step cost is similar to explicit Euler, and no nonlinear system is solved at each step.
In high-dimensional or mean-field settings, taming is localized per particle or component, enabling parallelism.

5. Current Frontiers and Open Problems

Sharpness of Strong Rates: For non-globally Lipschitz drift, the achievable strong rate remains $1/2$ without additional smoothness/structure; improved rates for smoother settings (e.g., order $1$) require further regularity or specialized weak error analysis (e.g., for additive noise (Liu et al., 2024)).
Optimality of Taming Function: The choice of taming function (divisor exponent, smooth cutoff, localized truncation) balances bias, stability, and computational efficiency. Modified schemes (e.g., (Ju et al., 13 Jul 2025)) allow the taming error to be reduced below the discretization error.
Preservation of Invariant Measure: Quantitative Wasserstein or total variation metrics allow for estimation of the stationary distribution approximation error; recent works provide explicit and uniform-in-time error bounds (Bao et al., 2024).
Extensions to Path-Dependent, Infinite-Dimensional Systems: Recent advances extend taming to stochastic evolution equations, delayed systems, neutral systems, and SPDEs, but explicit strong rates in the SPDE context remain challenging (Gyöngy et al., 2014).
Coupling and Ergodicity: Reflection coupling and Lyapunov techniques underpin recent results on geometric ergodicity for both the exact and numerical invariant measure (Bao et al., 2024, Liu et al., 2024).
Nonclassical Regimes: Taming can address SDEs with non-Locally integrable or truly singular drift, a regime intractable with standard explicit or even implicit methods (Johnston et al., 2024, Johnston et al., 23 Jan 2026).

6. Applications and Numerical Evidence

Recent literature confirms the theoretical rates and uniform moment bounds for tamed Euler schemes via:

SGLD with superlinear drift (tamed SGLD) achieving precise long-time KL divergence bounds between numerical and true stationary laws (Ju et al., 13 Jul 2025).
Mean-field (McKean–Vlasov) models in finance, statistical mechanics, and interacting particles, attaining $O(1/2)$ convergence, exponential ergodicity, and sharp propagation of chaos (Soni et al., 18 Oct 2025, Kumar et al., 2020).
Lévy-driven and jump SDEs, including strong convergence for discontinuous or merely measurable drift coefficients (Biswas et al., 21 Oct 2025, Neelima et al., 2020).
Stochastic transport and PDEs with singular coefficients, for which classical schemes fail (Lê et al., 2021, Johnston et al., 23 Jan 2026).

7. Comparison with Alternative and Classical Methods

Explicit Euler–Maruyama diverges or fails for superlinear (non-globally Lipschitz) drift (Sabanis, 2013).
Implicit/Backward Euler is stable but computationally expensive (requires solving nonlinear equations at every step).
Truncated Schemes (projection onto compact sets) restore moment stability but may introduce significant bias and require stopping times or event-based calculus, complicating error analysis (Ju et al., 13 Jul 2025).
Moreau–Yosida Regularization preserves monotonicity but is implicit and requires optimization at each step (Johnston et al., 2021).
Tamed Euler Schemes are fully explicit, require only evaluation of coefficients and bounded divisions, extendable to non-globally Lipschitz, singular, and infinite-dimensional settings, and match the strong order of classical Euler methods in their regime of applicability.