A modified tamed scheme for stochastic differential equations with superlinear drifts

Lei Li; Yuliang Wang; Zichang Ju

REVIEW 2 major objections 3 minor 1 cited by

A modified tamed scheme for stochastic differential equations with superlinear drifts

T0 review · 2 major / 3 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Taming only the large drift values keeps Euler-type SDE schemes stable at their original strong and weak orders, and gives near-second-order long-time sampling error.

desk verdict Cutoff-taming is a clean, useful idea and the finite-time theory is solid; the long-time KL bound, however, only covers random-batch models with uniformly bounded batch errors, which misses the standard superlinear SGLD setting. read the letter →

arxiv 2507.09475 v1 pith:RSZBSU7A submitted 2025-07-13 math.NA cs.NA

Zichang Ju , Lei Li , Yuliang Wang This is my paper · ORCID

classification math.NAcs.NA MSC 65C3060H3565C0560J60

keywords modifiedtamedschemeEuler-MaruyamasuperlineardriftstochasticgradientLangevindynamicsuniform-in-timeerrorestimaterelativeentropyrandombatchapproximationweakconvergence

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Explicit Euler-type discretizations blow up when SDE drifts grow superlinearly, and existing tamed or truncated fixes either lose accuracy or force awkward stopping-time analysis. This paper proposes a modified tamed scheme that switches on the taming only when the drift is large, via a smooth cutoff, so that the numerical drift equals the true drift on typical values. The paper proves that this modified tamed Euler scheme recovers the strong order 1/2 and weak order 1 of the underlying Euler discretization, matching the globally Lipschitz regime, even when the drift is replaced by a random-batch estimator. Applied to stochastic gradient Langevin dynamics, it yields a tamed SGLD algorithm with a uniform-in-time relative entropy bound of order $h^{2-\epsilon}$, a near-sharp analogue of the sharp $h^2$ rate known for Lipschitz drifts.

What carries the argument

The modified tamed drift $b^h(t,x)=b(t,x)/(1+\psi(\gamma h^\alpha |b(t,x)|))$, with cutoff $\psi(r)=0$ for $r\le 1$ and $\psi(r)=r$ for $r\ge 2$, is the central object. It retains the usual tamed bound $|b^h|\le \min(2\gamma^{-1}h^{-\alpha}, |b|)$, so moments stay bounded, while the discrepancy $b-b^h$ is nonzero only on the event $|b|\ge \gamma^{-1}h^{-\alpha}$. Lemma 3.5 converts the small probability of that event into an $O(h^q)$ taming error for any $q$, and this error bound is what lets the proofs preserve the Euler scheme's strong and weak orders and produce the $h^{2-\epsilon}$ relative entropy rate.

What would settle it

Take a two-batch estimator in which one batch gradient equals $-\nabla U(x)$ plus a term growing like $|x|^{\ell+1}$ on a set of positive measure, violating $\sup_\xi\sup_x|b_\xi(x)-b(x)|<\infty$, or a batch satisfying $x\cdot b_\xi(x)>0$ for arbitrarily large $|x|$, violating the angle condition. Then the random-batch term in Step 4 of the proof of Theorem 4.1 no longer admits a uniform-in-time bound, and a simulation of T-SGLD on $U(x)=|x|^4/4$ with such a batch should show relative entropy growing with time instead of staying $O(h^{2-\epsilon})$.

Watch

Extended reading notes

Core claim

The central claim is that taming only far from the origin, by replacing $b$ with $b^h = b/(1+\psi(h^\alpha|b|))$ where $\psi$ vanishes for arguments at most 1, preserves the accuracy of explicit discretizations for superlinear drifts while keeping them stable. Theorem 3.1 establishes $\sup_{0\le t\le T} E|X(t)-\hat X^h(t)|^2 \le C h$, and Theorem 3.2 establishes first-order weak accuracy, under one-sided Lipschitz and polynomial-growth assumptions. Theorem 4.1 establishes $\sup_{t\ge 0} H(\bar\rho_t^h|\rho_t) \le C h^{2-\epsilon}$ for the tamed SGLD algorithm with taming parameter $\alpha = \epsilon/2$. The mechanism is an $L^p$ estimate of the taming error that is $O(h^q)$ for arbitrarily large $q$, because the numerical solution visits the tamed region only with probability that can be made arbitrarily small in $h$.

Load-bearing premise

For the long-time sampler bound, every minibatch drift must stay uniformly close to the true drift and uniformly push toward the origin at large $|x|$; if some batch can disagree unboundedly or push outward, the uniform-in-time KL estimate is not obtained.

Editorial extensions

If this is right

The modified tamed Euler scheme converges strongly with order 1/2 and weakly with order 1 for SDEs with superlinear drift, the same rates as for globally Lipschitz coefficients.
With additive noise and no random batch, the strong order becomes 1; with a random batch the strong order drops to 1/2 while the weak order stays 1.
The tamed SGLD algorithm samples from superlinear targets with a uniform-in-time relative entropy error of order $h^{2-\epsilon}$ for any $\epsilon>0$, so long runs do not accumulate extra discretization error.
Because the strategy is scheme-agnostic, applying the same cutoff to a Milstein discretization should give first-order strong and weak rates, removing the half-order weak limitation noted for standard tamed Milstein schemes.
Random-batch approximation of the drift does not degrade the convergence rates, so minibatching can cut computational cost without sacrificing the theoretical accuracy of the scheme.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The cutoff construction suggests a general recipe: apply a stabilizing regularization only where the state is rare, and use moment and Markov bounds to make the regularization error arbitrarily high order; the same recipe may transfer to tamed discretizations of stochastic PDEs or jump processes.
The $h^{2-\epsilon}$ gap to the sharp $h^2$ Lipschitz result appears tied to using the $h^{-\alpha}$ bound on $\nabla b^h$ in the conditional-expectation estimate; if a uniform-in-time bound on $\int |\nabla\log\bar\rho_t|^{2+\delta}\bar\rho_t\,dx$ can be established, an optimal second-order rate should follow, as the paper itself flags in Section 6.
A practical testable extension: on a Bayesian logistic regression with a heavy-tailed prior, minibatched T-SGLD should show relative entropy error scaling like $h^{2-\epsilon}$ uniformly in time, while ordinary SGLD diverges because the drift is superlinear.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

First, the good news. The cutoff-taming construction (2.2)-(2.3) is genuinely nice: tame only when |b| is huge, and you get the moment control of taming without the whole-line distortion that costs weak order. Lemma 3.5—the taming error can be pushed to arbitrarily high order in L^p—is the right trick and it pays off. The finite-time strong and weak error analyses (Theorems 3.1, 3.2) are careful and, as far as I can tell, correct; the one-step estimates in Lemma 3.7 are the kind of detail that makes or breaks this sort of paper, and they hold up. If you want to run an explicit Euler-type scheme on a superlinear drift and you need first-order weak accuracy, this is now the method I would reach for.

Now the soft spots. The stress-test note lands. Theorem 4.1 depends on Assumption 4.1(b): sup_xi,x |b_xi(x)-b(x)| < infinity. For the motivating finite-sum SGLD, b_xi is the minibatch average; unless every component b_i differs from b by a bounded function, this difference grows like |x| when the b_i have different superlinear parts—or even different linear slopes. So the uniform-in-time KL bound does not cover the advertised application, and the numerical example (which batches the linear part of a split drift) still has an unbounded batch error. This is an applicability gap, not an internal contradiction. The finite-time theorems use the weaker and more reasonable Assumption 3.1 and are not affected. The KL proof also leans heavily on [21]: Lemma 4.3 is a sketch, and B.3/B.4 are asserted. The overall structure is standard, but a referee would want those filled in or at least stated precisely. The numerics are consistent with the rates, but they use the same scheme for the reference solution, no error bars, no code; that's fine for a sanity check, not for independent confirmation.

Bottom line: this paper deserves a serious referee. The construction is worth publishing on its own, the finite-time analysis is solid, and the KL theorem is true for a restricted class. The authors should be asked to either generalize the random-batch assumption or be upfront that the KL bound applies only when the batch error is bounded, and to fill the gaps in the KL proof. I'd cite this for the MTE scheme; I'd read the long-time sampling result with that caveat.

Referee Report

2 major / 3 minor

Summary. The paper proposes a modified tamed Euler (MTE) scheme for SDEs with superlinearly growing drifts, introducing a cut-off function in the taming so that the drift is modified only for large |b(x)|. The scheme allows an inaccurate (random-batch) drift. The authors prove strong order 1/2 and weak order 1 for the MTE scheme under one-sided Lipschitz and polynomial-growth assumptions (Theorems 3.1 and 3.2), and then apply the scheme to stochastic gradient Langevin dynamics (T-SGLD), proving a uniform-in-time relative-entropy bound of order h^{2-ε} (Theorem 4.1). Numerical experiments on a 1D Ginzburg-Landau SDE and a 2D Langevin equation are reported.

Significance. The finite-time strong and weak convergence results are a genuine contribution: the cut-off taming idea is simple, the local taming error estimate in Lemma 3.5 is elegant and essentially removes the taming error from the convergence rate, and the proofs follow standard Itô-calculus arguments. If Theorem 4.1 is fully established, the uniform-in-time KL bound for T-SGLD with superlinear drift would be an important step beyond the globally Lipschitz setting. The paper is clearly written overall and the numerical experiments support the claimed orders. However, two load-bearing issues in the long-time SGLD part need to be addressed.

major comments (2)

[Assumptions 4.1 and Section 4, Theorem 4.1] Assumption 4.1(b)-(c) are not satisfied by the standard finite-sum minibatch construction described in Section 2. For b(x)=b0(x)+(1/N)Σ_i b_i(x) and b_ξ(x)=b0(x)+(1/|S|)Σ_{i∈S} b_i(x), the difference b_ξ-b is an average of a random subset minus the full average. If any individual component b_i is superlinear (e.g., b_i(x)=|x|^2 x for a quadratic term of the potential), then sup_x |b_ξ(x)-b(x)| is unbounded, violating Assumption 4.1(b). Likewise, Assumption 4.1(c) requires every batch estimator to be dissipative, which need not hold for heterogeneous minibatches. Since Lemma 4.1 (Lyapunov bound), Lemma 4.3, and Step 4 of the proof of Theorem 4.1 all rely on Assumption 4.1(b)-(c), Theorem 4.1 as stated does not cover the motivating T-SGLD setting with general superlinear finite-sum drifts. The numerical T-SGLD experiment in Section 5.2 also appears not to satisfy Assumption 4.1(b) if the superlinear part is minibatched. Please clarify the intended scope of Theorem 4.1 or add a discussion/example of a random-batch model that does satisfy these assumptions.
[Section 4.2, Step 5 (after Eq. (4.19))] The transition from ∫ e^{-C2(t-s)} I(ρ_{κ(s)}) ds to a bound involving ∫ e^{-C2(t-s)} I(ρ_s) ds is not justified by Lemma 4.4 as stated. Lemma 4.4 proves a forward bound of the form I(ρ_{T_{k-1}+Δt}) ≤ 4I(ρ_{T_{k-1}})+C(Δt)^2, which controls future Fisher information by the value at the left endpoint of the time step. It does not provide the reverse control of the grid-point value I(ρ_{κ(s)}) in terms of the continuous integral over the preceding interval. A spike of I(ρ_s) at a grid point would not be detected by the weighted integral of I(ρ_s), and the forward inequality alone cannot rule this out. The sentence claiming that Lemma 4.4 lets one control I(ρ_{κ(s)}) by I(ρ_s) for s∈[κ(s)-h,κ(s)) needs a complete argument, or a citation to the corresponding step in [21] with the adaptation to the present non-Lipschitz setting made explicit. Without this, the final Gronwall step of Theorem 4.1 is incomplete.

minor comments (3)

[Lemma 3.7(c)] In the proof of Lemma 3.7(c), Y(t) is defined as (X̂(t)-X(T_k))·Z'·(X(t)-X̂(T_{k+1})), but the displayed Itô expansion and the three terms in (3.17) correspond to Y(t)=(X̂(t)-X(T_k))·Z'·(X(t)-X̂(t)). Please correct the definition of Y(t) or the subsequent expansion.
[Appendix B.1 and Lemmas B.3-B.4] Lemma 4.3 is only sketched in Appendix B.1, and Lemmas B.3-B.4 are asserted with very brief justifications. Since these lemmas are load-bearing for the uniform-in-time KL estimate, please expand the derivations, in particular the use of Jensen's inequality and Itô's formula in B.3 and B.4.
[Throughout] There are numerous typos and minor wording issues, e.g., 'supelinear', 'albe', 'nongloabally', 'soluitons', 'Motovated', and inconsistent notation such as Tk vs T_{k-1} in Section 4.2. A careful proofreading pass is recommended.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the strong, weak, and relative-entropy error bounds are derived from explicit estimates, not from the target rates or from fitted parameters.

full rationale

The derivation chain is self-contained. Theorem 3.1 splits the strong error into five terms (Eq. 3.8); the taming term I5 is controlled by Lemma 3.5, where the taming error is bounded by Markov's inequality with an arbitrarily high power of h, so the 1/2 order is not assumed. Theorem 3.2 uses one-step estimates (Lemma 3.7) that give O(h^2) per step, yielding O(h) globally. Theorem 4.1 obtains d/dt H <= C1 h^{2(1-alpha)}(1+I) - C2 H from the explicit J1, J2, J3 estimates in Steps 2-4, with the uniform Fisher-information bounds supplied by Lemmas 4.3 and 4.4; setting alpha = epsilon/2 then gives h^{2-epsilon}. The heavy use of the authors' prior paper [21] is methodological: the estimates are re-derived or adapted to the superlinear tamed setting, and [21] treats Lipschitz SGLD, so it is not a definitional input. Assumption 4.1(b)-(c) restrict the admissible random-batch estimators; this is an applicability limitation rather than circularity. The numerical reference being the same MTE scheme at a finer step is a validation caveat, but it does not enter any theorem proof.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

The central theorems add no fitted constants. The adjustable design parameters alpha and gamma are arbitrary within their allowed ranges, so they are not data-fitted. The real burden sits in the imported strong-dissipativity and bounded-batch-error assumptions, together with the relative-entropy technology inherited from Li and Wang (2021).

free parameters (2)

Taming exponent alpha = alpha in (0,1/2]; numerics use alpha=1/2; T-SGLD rate sets alpha=epsilon/2
Controls how large |b| must be before taming activates. The final KL rate depends on alpha, but no value is fitted to data.
Cutoff scale gamma = set to 1 in theory; gamma=1 and gamma=0.1 in numerical examples
Scales the taming activation threshold. The theory fixes gamma=1 and does not estimate it from data.

assumptions (6)

domain assumption Assumption 3.1: one-sided Lipschitz, polynomial growth, and time regularity for b_xi and sigma, uniformly in the batch variable xi.
Used throughout Section 3 to prove moment bounds, strong error, and weak error for the MTE scheme with random batch.
domain assumption Assumption 3.3: nondegenerate elliptic sigma sigma^T, bounded derivatives of Lambda, and polynomial derivative growth for b.
Needed for polynomial derivative bounds of the backward Kolmogorov solution in Lemma 3.6, hence for the first-order weak expansion in Theorem 3.2.
domain assumption Assumption 4.1(b)-(c): sup_xi sup_x |b_xi(x)-b(x)| < infinity and uniform dissipativity of each batch drift for |x| > R.
Load-bearing for the uniform-in-time Lyapunov bound in Lemma 4.1 and for the random-batch entropy estimate in Step 4 of Theorem 4.1.
domain assumption Assumption 4.2: the invariant measure pi satisfies a log-Sobolev inequality and rho_0/pi is bounded with finite exponential moment.
Provides the uniform LSI for rho_t and initial integrability required by the relative-entropy proof in Section 4.
domain assumption The sharp SGLD relative-entropy framework of Li and Wang (2021), including their Lemmas 2.1, 3.1, 3.2 and Appendix A.3, is imported.
Theorem 4.1 explicitly says parts of the proof are very similar to that paper, and several supporting lemmas here are only sketched or omitted.
standard math Standard stochastic calculus facts: Itô formula, Burkholder-Davis-Gundy inequality, Grönwall inequality, Stam convolution inequality, Girsanov theorem, and Holley-Stroock perturbation.
Used without proof throughout Sections 3 and 4 and in the appendices.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A modified tamed scheme for stochastic differential equations with superlinear drifts." pith.science (2026). https://pith.science/paper/RSZBSU7A

@misc{pith2026250709475,
  author       = {Pith},
  title        = {Pith review of: A modified tamed scheme for stochastic differential equations with superlinear drifts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RSZBSU7A}},
  note         = {Machine review of arXiv:2507.09475}
}

read the original abstract

Explicit discretizations of stochastic differential equations often encounter instability when the coefficients are not globally Lipschitz. The truncated schemes and tamed schemes have been proposed to handle this difficulty, but truncated schemes involve analyzing of the stopping times while the tamed schemes suffer from the reduced order of accuracy. We propose a modified tamed scheme by introducing an additional cut-off function in the taming, which enjoys the convenience for error analysis and preserving the original order of explicit discretization. While the strategy could be applied to any explicit discretization, we perform rigorous analysis of the modified tamed scheme for the Euler discretization as an example. Then, we apply the modified tamed scheme to the stochastic gradient Langevin dynamics for sampling with super-linear drift, and obtain a uniform-in-time near-sharp error estimate under relative entropy.

Figures

Figures reproduced from arXiv: 2507.09475 by the authors.

**Figure 2.** Weak convergence for the 1D example. Both the TME with and without random [PITH_FULL_IMAGE:figures/full_fig_p024_2.png] view at source ↗

**Figure 3.** Strong convergence for the 2D Langevin equation with additive noise. The MTE [PITH_FULL_IMAGE:figures/full_fig_p025_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Weak convergence for the 2D Langevin equation. (Test function, Left: exp( [PITH_FULL_IMAGE:figures/full_fig_p025_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

RELTA-SGLD: Relative-Growth Localized Taming for Nonconvex Stochastic-Gradient Langevin Learning
stat.ML 2026-07 conditional novelty 6.0 of 10

A relative-growth, threshold-localized taming denominator for SGLD achieves O(λ) stationary W1 and (in the potential case) W2 accuracy for nonconvex superlinear stochastic-gradient oracles.

Reference graph

Works this paper leans on

41 extracted references · 36 canonical work pages · cited by 1 Pith paper

[21]

A sharp uniform-in-time error estimate for Stochastic Gra- dient Langevin Dynamics

Lei Li and Yuliang Wang. A sharp uniform-in-time error estimate for Stochastic Gra- dient Langevin Dynamics. CSIAM Transactions on Applied Mathematics , 2025

work page 2025
[1]

Analysis and geometry of Markov diffusion operators, volume 348

Dominique Bakry, Ivan Gentil, and Michel Ledoux. Analysis and geometry of Markov diffusion operators, volume 348. Springer Science & Business Media, 2013

work page 2013
[2]

Sur la g´ en´ eralisation du probl` eme de Dirichlet: Premi` ere partie.Math- ematische Annalen, 62(2):253–271, 1906

Serge Bernstein. Sur la g´ en´ eralisation du probl` eme de Dirichlet: Premi` ere partie.Math- ematische Annalen, 62(2):253–271, 1906

work page 1906
[3]

Sur la g´ en´ eralisation du probl` eme de Dirichlet: Deuxi` eme partie

Serge Bernstein. Sur la g´ en´ eralisation du probl` eme de Dirichlet: Deuxi` eme partie. Mathematische Annalen, 69(1):82–136, 1910

work page 1910
[4]

On the weak convergence rate of an exponential Euler scheme for SDEs governed by coefficients with superlinear growth

Mireille Bossy, Jean-Fran¸ cois Jabir, and Kerlyns Mart ´ ınez. On the weak convergence rate of an exponential Euler scheme for SDEs governed by coefficients with superlinear growth. 2021

work page 2021
[5]

A maximal function characterization of the class H p

Donald L Burkholder, Richard F Gundy, and Martin L Silverstein. A maximal function characterization of the class H p. Transactions of the American Mathematical Society , 157:137–153, 1971

work page 1971
[6]

The truncated EM method for stochastic differential equations with Poisson jumps

Shounian Deng, Weiyin Fei, Wei Liu, and Xuerong Mao. The truncated EM method for stochastic differential equations with Poisson jumps. Journal of Computational and Applied Mathematics, 355:232–257, 2019

work page 2019
[7]

A collision-oriented interacting particle system for Landau-type equations and the molecular chaos

Kai Du and Lei Li. A collision-oriented interacting particle system for Landau-type equations and the molecular chaos. arXiv preprint arXiv:2408.16252 , 2024

arXiv 2024

Show all 41 references

[8]

Adaptive Euler–Maruyama method for SDEs with nonglobally Lipschitz drift

Wei Fang and Michael B Giles. Adaptive Euler–Maruyama method for SDEs with nonglobally Lipschitz drift. The Annals of Applied Probability , 30(2):526–560, 2020

2020
[9]

Propagation of chaos for the 2D viscous vortex model

Nicolas Fournier, Maxime Hauray, and St´ ephane Mischler. Propagation of chaos for the 2D viscous vortex model. Journal of the European Mathematical Society , 16(7):1423– 1466, 2014

2014
[10]

The theory of stochastic processes II

Iosif Ilich Gikhman and Anatoliui Vladimirovich Skorokhod. The theory of stochastic processes II. Springer Science & Business Media, 2004

2004
[11]

Convergence analysis of an explicit method and its random batch approximation for the McKean–Vlasov equations with non-globally Lipschitz conditions

Qian Guo, Jie He, and Lei Li. Convergence analysis of an explicit method and its random batch approximation for the McKean–Vlasov equations with non-globally Lipschitz conditions. ESAIM: Mathematical Modelling and Numerical Analysis , 58(2):639–671, 2024

2024
[12]

The partially truncated Euler– Maruyama method and its stability and boundedness

Qian Guo, Wei Liu, Xuerong Mao, and Rongxian Yue. The partially truncated Euler– Maruyama method and its stability and boundedness. Applied Numerical Mathematics, 115:235–251, 2017. 33

2017
[13]

Mean field error estimate of the random batch method for large interacting particle system

Zhenyu Huang, Shi Jin, and Lei Li. Mean field error estimate of the random batch method for large interacting particle system. ESAIM: Mathematical Modelling and Numerical Analysis, 59(1):265–289, 2025

2025
[14]

Strong and weak divergence in finite time of euler’s method for stochastic differential equations with non-globally Lipschitz continuous coefficients

Martin Hutzenthaler, Arnulf Jentzen, and Peter E Kloeden. Strong and weak divergence in finite time of euler’s method for stochastic differential equations with non-globally Lipschitz continuous coefficients. Proceedings of the Royal Society A: Mathematical, Physical and Engin...

2011
[15]

Strong convergence of an explicit numerical method for SDEs with nonglobally Lipschitz continuous coefficients

Martin Hutzenthaler, Arnulf Jentzen, and Peter E Kloeden. Strong convergence of an explicit numerical method for SDEs with nonglobally Lipschitz continuous coefficients. The Annals of Applied Probability , 22(4):1611–1641, 2012

2012
[16]

Random batch methods (RBM) for interacting particle systems

Shi Jin, Lei Li, and Jian-Guo Liu. Random batch methods (RBM) for interacting particle systems. Journal of Computational Physics , 400:108877, 2020

2020
[17]

Convergence of the random batch method for interacting particles with disparate species and weights

Shi Jin, Lei Li, and Jian-Guo Liu. Convergence of the random batch method for interacting particles with disparate species and weights. SIAM Journal on Numerical Analysis, 59(2):746–768, 2021

2021
[18]

Stochastic differential equations

Peter E Kloeden and Eckhard Platen. Stochastic differential equations. Springer, 1992

1992
[19]

On tamed Milstein schemes of SDEs driven by L´ evy noise.Discrete and Continuous Dynamical Systems-B , 22(2):421–463, 2016

Chaman Kumar and Sotirios Sabanis. On tamed Milstein schemes of SDEs driven by L´ evy noise.Discrete and Continuous Dynamical Systems-B , 22(2):421–463, 2016

2016
[20]

Error estimates of the Euler’s method for stochastic differential equations with multiplicative noise via relative entropy

Lei Li, Mengchao Wang, and Yuliang Wang. Error estimates of the Euler’s method for stochastic differential equations with multiplicative noise via relative entropy. arXiv preprint arXiv:2409.04991, 2024

2024 arXiv
[22]

Truncated Euler-Maruyama method for classical and time-changed non-autonomous stochastic differential equa- tions

Wei Liu, Xuerong Mao, Jingwen Tang, and Yue Wu. Truncated Euler-Maruyama method for classical and time-changed non-autonomous stochastic differential equa- tions. Applied Numerical Mathematics , 153:66–81, 2020

2020
[23]

Geometric ergodicity and optimal error estimates for a class of novel tamed schemes to super-linear stochastic pdes

Zhihui Liu and Jie Shen. Geometric ergodicity and optimal error estimates for a class of novel tamed schemes to super-linear stochastic pdes. arXiv preprint arXiv:2502.19117 , 2025

2025 arXiv
[24]

The truncated Euler–Maruyama method for stochastic differential equa- tions

Xuerong Mao. The truncated Euler–Maruyama method for stochastic differential equa- tions. Journal of Computational and Applied Mathematics , 290:370–384, 2015

2015
[25]

Convergence rates of the truncated Euler–Maruyama method for stochas- tic differential equations

Xuerong Mao. Convergence rates of the truncated Euler–Maruyama method for stochas- tic differential equations. Journal of Computational and Applied Mathematics , 296:362– 375, 2016

2016
[26]

Ergodicity for SDEs and approximations: locally Lipschitz vector fields and degenerate noise

Jonathan C Mattingly, Andrew M Stuart, and Desmond J Higham. Ergodicity for SDEs and approximations: locally Lipschitz vector fields and degenerate noise. Stochastic processes and their applications , 101(2):185–232, 2002

2002
[27]

Approximate integration of stochastic differential equations

Grigori N Milstein. Approximate integration of stochastic differential equations. Theory of Probability & Its Applications , 19(3):557–562, 1975

1975
[28]

Stochastic numerics for mathematical physics, volume 39

Grigori N Milstein and Michael V Tretyakov. Stochastic numerics for mathematical physics, volume 39. Springer, 2004

2004
[29]

Numerical integration of stochastic dif- ferential equations with nonglobally Lipschitz coefficients

Grigori N Milstein and Michael V Tretyakov. Numerical integration of stochastic dif- ferential equations with nonglobally Lipschitz coefficients. SIAM journal on numerical analysis, 43(3):1139–1154, 2005. 34

2005
[30]

Im- proved bounds for discretization of Langevin diffusions: Near-optimal rates without convexity

Wenlong Mou, Nicolas Flammarion, Martin J Wainwright, and Peter L Bartlett. Im- proved bounds for discretization of Langevin diffusions: Near-optimal rates without convexity. Bernoulli, 28(3):1577–1601, 2022

2022
[31]

Generalization bounds of sgld for non-convex learning: Two theoretical viewpoints

Wenlong Mou, Liwei Wang, Xiyu Zhai, and Kai Zheng. Generalization bounds of sgld for non-convex learning: Two theoretical viewpoints. In Conference on Learning Theory, pages 605–638. PMLR, 2018

2018
[32]

A stochastic approximation method

Herbert Robbins and Sutton Monro. A stochastic approximation method. The annals of mathematical statistics , pages 400–407, 1951

1951
[33]

Diffusions, Markov processes, and martingales: Itˆ o calculus, volume 2

L Chris G Rogers and David Williams. Diffusions, Markov processes, and martingales: Itˆ o calculus, volume 2. Cambridge university press, 2000

2000
[34]

A note on tamed euler approximations

Sotirios Sabanis. A note on tamed euler approximations. Electronic Communications in Probability, 18(a):47, 2013

2013
[35]

Euler approximations with varying coefficients: the case of super- linearly growing diffusion coefficients

Sotirios Sabanis. Euler approximations with varying coefficients: the case of super- linearly growing diffusion coefficients. Annals of Applied Probability , 26(4):2083–2105, 2016

2016
[36]

Some inequalities satisfied by the quantities of information of Fisher and Shannon

Aart J Stam. Some inequalities satisfied by the quantities of information of Fisher and Shannon. Information and Control , 2(2):101–112, 1959

1959
[37]

Consistency and fluctuations for stochastic gradient Langevin dynamics

Yee W Teh, Alexandre H Thiery, and Sebastian J Vollmer. Consistency and fluctuations for stochastic gradient Langevin dynamics. The Journal of Machine Learning Research, 17(1):193–225, 2016

2016
[38]

The tamed Milstein method for commutative stochastic differential equations with non-globally Lipschitz continuous coefficients

Xiaojie Wang and Siqing Gan. The tamed Milstein method for commutative stochastic differential equations with non-globally Lipschitz continuous coefficients. Journal of Difference Equations and Applications , 19(3):466–490, 2013

2013
[39]

Weak error analysis for strong ap- proximation schemes of SDEs with super-linear coefficients

Xiaojie Wang, Yuying Zhao, and Zhongqiang Zhang. Weak error analysis for strong ap- proximation schemes of SDEs with super-linear coefficients. IMA Journal of Numerical Analysis, 44(5):3153–3185, 2024

2024
[40]

Random batch method for stochastic differential equations with non- globally Lipschitz coefficients

Xuebing Wang. Random batch method for stochastic differential equations with non- globally Lipschitz coefficients. Master’s thesis, (in Chinese), Shanghai Jiao Tong Uni- versity, 2024

2024
[41]

Bayesian learning via stochastic gradient Langevin dynamics

Max Welling and Yee W Teh. Bayesian learning via stochastic gradient Langevin dynamics. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 681–688. Citeseer, 2011. 35

2011

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [21]

A sharp uniform-in-time error estimate for Stochastic Gra- dient Langevin Dynamics

Lei Li and Yuliang Wang. A sharp uniform-in-time error estimate for Stochastic Gra- dient Langevin Dynamics. CSIAM Transactions on Applied Mathematics , 2025

work page 2025

[2] [1]

Analysis and geometry of Markov diffusion operators, volume 348

Dominique Bakry, Ivan Gentil, and Michel Ledoux. Analysis and geometry of Markov diffusion operators, volume 348. Springer Science & Business Media, 2013

work page 2013

[3] [2]

Sur la g´ en´ eralisation du probl` eme de Dirichlet: Premi` ere partie.Math- ematische Annalen, 62(2):253–271, 1906

Serge Bernstein. Sur la g´ en´ eralisation du probl` eme de Dirichlet: Premi` ere partie.Math- ematische Annalen, 62(2):253–271, 1906

work page 1906

[4] [3]

Sur la g´ en´ eralisation du probl` eme de Dirichlet: Deuxi` eme partie

Serge Bernstein. Sur la g´ en´ eralisation du probl` eme de Dirichlet: Deuxi` eme partie. Mathematische Annalen, 69(1):82–136, 1910

work page 1910

[5] [4]

On the weak convergence rate of an exponential Euler scheme for SDEs governed by coefficients with superlinear growth

Mireille Bossy, Jean-Fran¸ cois Jabir, and Kerlyns Mart ´ ınez. On the weak convergence rate of an exponential Euler scheme for SDEs governed by coefficients with superlinear growth. 2021

work page 2021

[6] [5]

A maximal function characterization of the class H p

Donald L Burkholder, Richard F Gundy, and Martin L Silverstein. A maximal function characterization of the class H p. Transactions of the American Mathematical Society , 157:137–153, 1971

work page 1971

[7] [6]

The truncated EM method for stochastic differential equations with Poisson jumps

Shounian Deng, Weiyin Fei, Wei Liu, and Xuerong Mao. The truncated EM method for stochastic differential equations with Poisson jumps. Journal of Computational and Applied Mathematics, 355:232–257, 2019

work page 2019

[8] [7]

A collision-oriented interacting particle system for Landau-type equations and the molecular chaos

Kai Du and Lei Li. A collision-oriented interacting particle system for Landau-type equations and the molecular chaos. arXiv preprint arXiv:2408.16252 , 2024

arXiv 2024