An explicit splitting SAV scheme for the kinetic Langevin dynamics

Lei Dai; Xiaojie Wang; Yingsong Jiang

REVIEW 4 major objections 3 minor 1 cited by

An explicit splitting SAV scheme for the kinetic Langevin dynamics

T0 review · 4 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper proves that an explicit splitting scalar auxiliary variable (SSAV) scheme for kinetic Langevin dynamics with possibly superlinear, non-globally Lipschitz potentials attains order-one strong convergence and order-one weak convergen

desk verdict A genuinely new explicit SAV scheme for kinetic Langevin with order-one strong and polynomial-in-time weak error; the construction and main proofs hold up, but Theorem 5.6 rests on a thinly verified, slightly mismatched ergodicity assumption. read the letter →

arxiv 2509.04251 v1 pith:FTV4FFKU submitted 2025-09-04 math.NA cs.NA

Lei Dai , Yingsong Jiang , Xiaojie Wang This is my paper

classification math.NAcs.NA MSC 60H3565C30

keywords kineticLangevindynamicssplittingscalarauxiliaryvariable(SSAV)schemeexplicitnumericalenergy-preservingintegrationexponentialintegrabilityorder-onestrongconvergencelong-timeweakinvariantmeasuresampling

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

The paper tries to establish that one can evolve kinetic Langevin dynamics—the underdamped stochastic system used in molecular dynamics and Hamiltonian Monte Carlo—by a fully explicit time-stepping scheme that nevertheless converges at the optimal rate, even when the potential's gradient grows superlinearly. The central target is a long-time weak error bound (Theorem 5.6): after a transient term that decays exponentially, the error in approximating an observable's expectation under the invariant measure is bounded by a constant times the stepsize h, where that constant grows polynomially, not exponentially, with the simulation time. The machinery is a splitting scalar auxiliary variable (SSAV) integrator: a scalar variable ρ is introduced to rewrite the energy as a sum of squares, the Hamiltonian part is integrated by an energy-preserving but apparently implicit step that turns out to be solvable in closed form, and the noise/friction part is integrated exactly. A polynomial-in-time growth of moments plus exponential integrability supplies the missing control, so the scheme stays practical for sampling high-dimensional targets.

What carries the argument

The central object is the scalar auxiliary variable ρ(t)=√(κΦ(u(t))+C_H−α|u(t)|²), where a quadratic term −α|u|² is deliberately subtracted from the potential before taking the square root (Remark 2.5). This turns the total energy into a sum of squares, H=|v|²/2+α|u|²+ρ², and makes the apparently implicit midpoint-type Hamiltonian step (2.10) algebraically solvable, so the whole scheme (2.10)–(2.11) is explicit (Proposition 2.4). The same ρ makes the scheme energy-preserving in the substep (Theorem 3.1) and, through an exponential-integrability lemma applied on each Ornstein–Uhlenbeck substep (Theorem 3.6), supplies uniform control of exp(η|u|²) and of higher moments. Strong convergence is t

What would settle it

Fix a potential from Example 2 (Φ(u)=|u|^4/4−|u|^2/2, m=20), a setting the paper's assumptions cover. Compute the strong error sup_{n≤N_T} E|X(t_n)-Y_{t_n}| for h=2^{-k}, k=6,...,11, at T=1 relative to a fine reference h=2^{-14}: the claimed Theorem 4.4 gives slope 1; a slope of 1/2 would contradict the optimal rate. Separately, measure |E[φ(Y^{x0}_{t_N})]−∫φdμ∞| at fixed h=2^{-9} for t_N=0,1,...,30: if, after the exponential transient, the error grows like e^{L t_N} h rather than staying within (1+t_N^l)h, then the polynomial time dependence of Theorem 5.6 is false.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that the SSAV scheme (2.10)–(2.11), for kinetic Langevin dynamics d v = −κ∇Φ(u)dt − γv dt + Γ dW_t, d u = v dt, has three coupled properties. First, it preserves a modified total energy exactly on each Hamiltonian substep, and the numerical energy grows in expectation only linearly in time. Second, over a finite interval [0,T] the numerical solution converges strongly with order one: E[sup_{n≤N_T}|X(t_n)-Y_{t_n}|^θ] ≤ C_T h^θ, with C_T independent of h (Theorem 4.4). Third, for exponentially ergodic dynamics satisfying a dissipativity/Lyapunov condition (Assumptions 5.2, 5.4), the long-time weak error satisfies |E[φ(Y^{x0}_{t_N})] - ∫φ dμ∞| ≤ C1 e^{-λt_N} +

Load-bearing premise

The long-time weak-error conclusion relies on the exact kinetic Langevin dynamics being exponentially ergodic with exponentially decaying first- and second-derivative semigroup bounds (Assumptions 5.2 and 5.4, via Lemmas 5.3 and 5.5); if a potential fails that dissipativity, the polynomial-in-time error estimate is not guaranteed.

Editorial extensions

If this is right

For target densities π∝exp(−Φ) with superquadratic, possibly nonconvex Φ, one can sample with an explicit integrator and still get first-order accuracy, avoiding the per-step cost of implicit methods in high dimension.
To reach total weak error ε in approximating ∫φ dμ∞, the scheme needs t_N=O(log(1/ε)) and h=O(ε(log(1/ε))^{-l}), so only O(ε^{-1}(log(1/ε))^{1+l}) iterations—polynomial in ε^{-1}, not exponential in a Lyapunov exponent.
The energy evolution bound (Corollary 3.2) means the numerical total energy stays controlled over long runs, which is exactly what prevents blow-up in sampling superlinear potentials.
Order-one strong convergence over finite [0,T] gives a computable, non-asymptotic error estimate for pathwise observables of kinetic Langevin dynamics with non-globally monotone drift.
The comparison with an implicit splitting scheme suggests that giving up exact uniform-in-time moment bounds costs only a logarithmic factor in the iteration count, while the explicit form saves substantially in high dimensions.

Reading between the lines

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

The same subtraction trick for ρ might be reusable for other SDEs/SPDEs whose energy is a polynomial plus a nonnegative perturbation, transferring exponential integrability to the numerical position variable without global Lipschitz assumptions.
One plausible next test: use the SSAV scheme as the sampler in stochastic-gradient Hamiltonian Monte Carlo; the energy-preserving substep may reduce momentum refreshment bias at large stepsize, though the paper does not analyze that.
The theorem's polynomial-in-time constant C2(1+t_N^l) likely hides values of l tied to moments of the Lyapunov/semigroup derivatives; identifying l explicitly for common potentials (Gaussian mixture, double well) would give a ready-to-use step-size rule.
If a dissipative Lyapunov function with uniform-in-time moment bounds were found for the SSAV scheme (as for the implicit scheme the paper compares with), the weak bound would improve to l=0, reducing iterations to O(ε^{-1} log(1/ε)); the paper notes this as the uniform case but does not claim it.

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

The headline result is real, as far as I can tell. The shifted auxiliary variable ρ = sqrt(κΦ(u) + C_H − α|u|²) is a smart modification of the standard SAV trick — the α|u|² subtraction is what buys exponential integrability of the position variable, without which the strong-rate argument would collapse. The explicit solvability (Prop 2.4), the energy identity (Thm 3.1), and the exponential integrability theorem (Thm 3.6) are clean, verifiable pieces of algebra, and the order-one strong convergence proof (Thm 4.4) is serious and mostly self-contained. The weak error theorem (Thm 5.6) with polynomial-in-time rather than exponential constants is the right headline: nothing else explicit gets that for kinetic Langevin with superlinear potentials and no convexity. The paper is honest about its scaffolding, and the self-citation to [11] is a published perturbation framework, so the circularity burden is genuinely low.

Soft spots, in proportion. First, Prop 4.2 and Thm 4.4 have a real quantification glitch: θ appears in the assumption (αδ > e^{λT}κC_Φ T^θ) while the conclusion is "for any θ > 0", and the proof in (4.29) actually needs q from a Hölder split. Repairable, but a referee shouldn't have to do it.

Second, and this is the load-bearing weakness: Theorem 5.6 depends entirely on Lemma 5.5, imported from Talay [30] and never proved here, and the "verification" of Assumption 5.4 for the examples is one sentence in Section 6. The stress-test's factor-of-2 point is technically correct: Assumption 5.4 writes the generator with 1/2 Δ_v, but (1.1) has 1/2 tr(ΓΓ^T ∂_vv), and none of the three examples has Γ = I. The fix is probably routine — the extra diffusion term is constant on the quadratic Lyapunov functions the examples use — but the paper doesn't do it, and "it is not difficult to check" is doing real work. If Lemma 5.5 fails or needs modification, the polynomial-in-time weak error and the §5.3 cost comparison both collapse. This is the part I would send back for a written check at the right generator, not just an assertion.

Third: no code or data shipped, so the experiments, which look fine and support the claimed rates, are illustrative only. There's also a typo in the N_exp formula in §5.3.

Verdict: the central construction holds up and the main theorems are largely proven; the weak estimate is standard once you grant Lemma 5.5. This deserves a serious referee. I would accept conditionally, on the Prop 4.2 fix and a real verification of Assumption 5.4 for the examples — I'd want to see that before relying on the long-time result myself.

Referee Report

4 major / 3 minor

Summary. The paper introduces an explicit splitting scalar auxiliary variable (SSAV) scheme for kinetic Langevin dynamics (1.1) with potentials whose gradients may grow superlinearly. By introducing a modified energy variable rho(t) under Assumption 2.3, the scheme is shown to be explicitly solvable (Prop 2.4) and to preserve a modified energy (Thm 3.1). The authors prove exponential integrability (Thm 3.6), polynomial moment growth (Prop 5.1), order-one strong convergence over finite time (Thm 4.4), and a long-time weak error estimate with polynomial-in-time constants (Thm 5.6). The weak error bound is then used to compare computational costs for invariant-measure sampling. Numerical experiments on Gaussian mixture, double-well, and bimodal potentials illustrate the predicted rates and sampling stability.

Significance. If the central results are correct, the paper makes a valuable contribution: an explicit, energy-stable scheme for superlinearly growing kinetic Langevin dynamics with order-one strong convergence and a long-time weak error whose constant grows only polynomially in time. The algebraic parts of the paper are clean: Prop 2.4 gives explicit update formulas, Thm 3.1 is a neat energy identity, and Thm 3.6 gives exponential integrability by a direct application of the cited perturbation criterion. The potential practical gain over implicit integrators, especially in high dimensions, is clearly motivated. However, the long-time weak claim (Theorem 5.6) and part of the strong convergence proof depend on hypotheses and estimates that are not fully justified in the manuscript, as detailed below.

major comments (4)

[Section 4.1, Eqs. (4.15)-(4.19)] The proof of Proposition 4.2 appears to contain a circular bound. After the Gronwall step, the term B_{k,2} is defined with B_{k,2} = (kappa(nabla Phi(u_n)-nabla Phi(u(t_n))), v(t_n)-v_n) h, which contains the error e_k = X(t_k)-Y_{t_k}. Equation (4.19) asserts ||B_{k,2}||_{L^p}^2 <= C h^2 'owing to Corollary 3.7', but Corollary 3.7 only controls moments of u_n, u(t_n), v_n, v(t_n) separately; it does not control their difference. A bound of size h^2 on B_{k,2} would essentially already imply the half-order result and is not available at this stage. The term B_{k,2} must instead be absorbed into the Gronwall recurrence on the left side via Young's inequality, or an external perturbation estimate from [11] must be invoked explicitly. As written, the derivation of (4.30) is not justified.
[Section 5.2, Assumption 5.4 and Lemma 5.5] The imported derivative-decay estimate is stated for a generator that does not match (1.1). Assumption 5.4 defines L psi = <v, partial_u psi> - <kappa nabla Phi(u)+gamma v, partial_v psi> + (1/2) sum_i partial_{v_i v_i} psi, i.e. the diffusion part is (1/2)Delta_v. But the generator of (1.1) has diffusion part (1/2)tr(Gamma Gamma^* D_v^2 psi). For the paper's own examples one has Gamma=2, Gamma=sqrt(2)I_m, and Gamma=0.1, none of which gives (1/2)Delta_v. Lemma 5.5 is imported verbatim from [30], and Theorem 5.6 uses this lemma through (5.19)-(5.20) for eS1, eS2 and eS3. Unless a rescaling is stated and executed, or Lemma 5.5 is proved for the actual generator, the polynomial-in-time weak error bound rests on an unverified hypothesis. Section 6's one-line assertion that the examples satisfy Assumption 5.4 does not repair this mismatch.
[Section 6, verification of Assumptions 5.2 and 5.4] The claim that Examples 1-3 satisfy Assumptions 5.2 and 5.4 is not substantiated. For Assumption 5.4 the paper states only 'It is not difficult to check...' and refers to [30, Example 1.1, Example 1.2]. Since Lemma 5.5 is entirely imported and is load-bearing for Theorem 5.6, the referee cannot verify that the required Lyapunov function R, the constants vartheta, tilde C, r, and the decay rates eta_1, eta_2 exist for the potentials and Gamma matrices used in the experiments. At minimum, the manuscript should provide the explicit R (or a precise proof) and state the resulting constants, or clearly explain how [30]'s examples transfer to the present parameter regime.
[Section 4, Proposition 4.2 and Theorem 4.4] The hypothesis 'alpha delta > e^{lambda T} kappa C_Phi T theta' is not well posed. In the statement theta is an arbitrary positive number, but in the proof the analogous condition is 'alpha delta >= e^{lambda T} kappa C_Phi T q', where q is the Holder conjugate used in (4.14)-(4.15). The theorem as stated therefore does not specify the correct condition on delta, alpha, and T. This matters because the exponential integrability bound (3.17) is used with a rate depending on q, and the proof's validity depends on this choice. The statement should be corrected and the proof's notation reconciled.

minor comments (3)

[Section 5, Theorem 5.6] The theorem statement does not explicitly specify the admissible test functions phi. Section 5.1 fixes phi in C^2(R^{2m};R) with derivative polynomial growth, but Theorem 5.6 should state this assumption in its own statement for self-containedness.
[Section 6, Eq. (6.1)] In Example 1, the parameter sigma and iota are used in the definition of Phi; the numerical choices iota=1, sigma=1/2 are given later, but the invariant measure formula (6.2) is written for general sigma. This is clear enough, but the notation N_m/mathcal N in (2.21) could be made explicit (standard normal increment independent of the past).
[Section 5.3, Eqs. (5.32)-(5.35)] The cost comparison would benefit from a short explanation of what 'iteration number' counts per Newton step, since for implicit schemes the per-step cost can depend on the nonlinear solver; as written the comparison only counts time steps.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation found; Theorem 5.6 is conditional on external ergodicity/derivative-decay assumptions rather than on the numerical scheme's own conclusions.

full rationale

The SSAV scheme is an explicit construction whose energy preservation (Theorem 3.1) is an algebraic identity built into the discretization of the auxiliary variable rho, not a prediction derived from the target error estimate. The paper's central results are proved from stated assumptions: exponential integrability (Theorem 3.6) via Lemma 3.4 applied to the OU substep; finite-time strong convergence (Theorem 4.4) via explicit Ito estimates, Gronwall/Jensen arguments, and moment bounds from Corollary 3.7 and Proposition 5.1; and long-time weak convergence (Theorem 5.6) by decomposing the weak error into eS1+eS2+eS3 and combining the numerical error estimate (5.29) with external ergodicity results Lemma 5.3 ([22]) and derivative-decay Lemma 5.5 ([30]). No parameter is fitted to the target result and no step assumes the conclusion. The citation of the authors' prior perturbation framework [11] in Section 4.1 is contextual; the subsequent proof is carried out in the paper, so that self-citation is not load-bearing. Per the reviewing rule, I flag the support/verification gaps rather than circularity: Lemma 5.5 is imported without proof, Assumptions 5.2/5.4 are verified for the examples only by one-line assertions ('It is not difficult to check...'), and the operator L in Assumption 5.4 has diffusion term 1/2 Sum_i partial_{v_i v_i}, which for examples with Gamma=sqrt(2)I differs from the generator of (1.1) by a factor of two unless an unstated rescaling is supplied. These are correctness risks because Theorem 5.6 is conditional on those hypotheses, but they are not circularity: the hypotheses concern the exact dynamics and are not derived from the numerical scheme's own convergence claim.

Assumptions & free parameters 3 free parameters · 6 assumptions · 1 invented entities

The central claim rests on the shifted-potential lower bound (Assumption 2.3), which justifies rho, and on external ergodicity theorems for the exact dynamics. The scheme parameters alpha and C_H are hand-chosen, not fitted to any target, and they do not encode the convergence rates.

free parameters (3)

alpha = set to 1 in experiments
Hand-chosen parameter in rho and in the modified energy; required to satisfy Assumption 2.3 and the condition alpha delta > e^{lambda T} kappa C_Phi T theta in Proposition 4.2.
C_H = set to 1000 in experiments
Large constant shifting the potential so that kappa Phi + C_H - alpha |u|^2 >= 1 and rho is well-defined.
delta = not specified
Parameter in the initial exponential-integrability assumption E[exp(delta(...))] < infinity; its size controls the allowed alpha delta product.

assumptions (6)

domain assumption Assumption 2.1: Phi in C^2 and a unique strong solution exists for the kinetic Langevin dynamics.
Standard well-posedness input for Ito calculus.
ad hoc to paper Assumption 2.3: for sufficiently large C_H, kappa Phi(x) + C_H - alpha |x|^2 >= 1 for all x.
Enables the definition of the auxiliary variable rho; this is the key structural restriction on the potential.
domain assumption Assumption 4.1: nabla Phi in C^2 with |nabla Phi(x) - nabla Phi(y)| <= C_Phi (1 + |x|^2 + |y|^2) |x - y|.
Polynomial Lipschitz condition allowing superlinear growth while bounding the Hessian quadratically; used throughout the strong convergence proof.
domain assumption Assumption 5.2: dissipativity condition on Phi guaranteeing exponential ergodicity of the exact dynamics.
Imported from [22, Condition 3.1]; used in Lemma 5.3 for the long-time weak convergence analysis.
domain assumption Assumption 5.4: existence of a Lyapunov function R with dissipativity and polynomial growth of Phi and all its derivatives.
Imported from [30, Hypothesis 1.1]; yields the semigroup derivative decay in Lemma 5.5 needed for the weak error proof.
standard math Cox-Hutzenthaler-Jentzen exponential integrability criterion (Lemma 3.4).
External theorem used to derive exponential integrability of both exact and numerical solutions.

invented entities (1)

rho(t) = sqrt(kappa Phi(u(t)) + C_H - alpha |u(t)|^2)
purpose: New scalar auxiliary variable that rewrites the total energy as |v|^2/2 + alpha |u|^2 + rho^2 and enables an explicit energy-preserving splitting.
Mathematical construct defined from the potential; there is no external falsifiable handle, and its properties are established inside the paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of An explicit splitting SAV scheme for the kinetic Langevin dynamics." pith.science (2026). https://pith.science/paper/FTV4FFKU

@misc{pith2026250904251,
  author       = {Pith},
  title        = {Pith review of: An explicit splitting SAV scheme for the kinetic Langevin dynamics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FTV4FFKU}},
  note         = {Machine review of arXiv:2509.04251}
}

abstract

The kinetic Langevin dynamics finds diverse applications in various disciplines such as molecular dynamics and Hamiltonian Monte Carlo sampling. In this paper, a novel splitting scalar auxiliary variable (SSAV) scheme is proposed for the dynamics, where the gradient of the potential $U$ is possibly non-globally Lipschitz continuous with superlinear growth. As an explicit scheme, the SSAV method is efficient, robust and is able to reproduce the energy structure of the original dynamics. By an energy argument, the SSAV scheme is proved to possess an exponential integrability property, which is crucial to establishing the order-one strong convergence without the global monotonicity condition. Moreover, moments of the numerical approximations are shown to have polynomial growth with respect to the time length. This helps us to obtain weak error estimates of order one, with error constants polynomially (not exponentially) depending on the time length. Despite the obtained polynomial growth, the explicit scheme is shown to be computationally effective for the approximation of the invariant distribution of the dynamics with exponential ergodicity. Numerical experiments are presented to confirm the theoretical findings and to show the superiority of the algorithm in sampling.

Figures

Figures reproduced from arXiv: 2509.04251 by the authors.

**Figure 1.** Strong and weak errors of the SSAV method in finite time horizon (a) Gaussian mixture potential with m = 1 (b) Double-well potential with m = 1 [PITH_FULL_IMAGE:figures/full_fig_p027_1.png] view at source ↗

**Figure 2.** Long-time weak errors of the SSAV method [PITH_FULL_IMAGE:figures/full_fig_p027_2.png] view at source ↗

**Figure 3.** Density approximation of Gaussian mixture potential with m = 1: (a) SSAV method with h = 2−4 and h = 2−7 ; (b) EM method with h = 2−2 ; (c) SSAV method with h = 2−2 27 [PITH_FULL_IMAGE:figures/full_fig_p027_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Density approximation of double-well potential with m = 2: (a) SSAV method with h = 2−4 ; (b) SSAV method with h = 2−7 ; (c) Standard probability density [PITH_FULL_IMAGE:figures/full_fig_p028_4.png]

**Figure 5.** Figure 5: Density approximation of bimodal distribution: (a) EM method with h = 2−4 ; (b) SSAV method with h = 2−4 ; (c) Standard probability density References [1] A. Abdulle, G. Vilmart, and K. C. Zygalakis. Long time accuracy of Lie–Trotter splitting methods for Langevin dyna…

Discussion (0). Sign in 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

Tamed Stochastic Gradient Hamiltonian Monte Carlo
math.OC 2026-07 conditional novelty 5.0 of 10

tSGHMC provably samples from strongly convex targets with superlinear, discontinuous stochastic gradients at a λ^{1/4} Wasserstein-2 rate.

Reference graph

Works this paper leans on

30 extracted references · 28 canonical work pages · cited by 1 Pith paper

[11]

Dai and X

L. Dai and X. Wang. Perturbation estimates for order-one strong approximations of SDEs without globally monotone coefficients. IMA Journal of Numerical Analysis , page draf034, 2025

work page 2025
[30]

D. Talay. Stochastic Hamiltonian systems: exponential convergence to the invariant measure, and discretization by the implicit Euler scheme.Markov Process. Related Fields, 8(2):163–198, 2002. 30

work page 2002
[1]

Abdulle, G

A. Abdulle, G. Vilmart, and K. C. Zygalakis. Long time accuracy of Lie–Trotter splitting methods for Langevin dynamics. SIAM Journal on Numerical Analysis , 53(1):1–16, 2015

work page 2015
[2]

C.-E. Brehier. Approximation of the invariant distribution for a class of ergodic SDEs with one-sided Lipschitz continuous drift coefficient using an explicit tamed Euler scheme.ESAIM: Probability and Statistics , 27:841–866, 2023

work page 2023
[3]

Y. Cao, J. Lu, and L. Wang. On explicit L2-convergence rate estimate for underdamped Langevin dynamics. Archive for Rational Mechanics and Analysis , 247(5):90, 2023

work page 2023
[4]

S. Cerrai. Second Order PDE’s in Finite and Infinite Dimension: a Probabilistic Approach . Springer, 2001

work page 2001
[5]

C. Chen, T. Dang, J. Hong, and F. Zhang. A new class of splitting methods that preserve ergodicity and exponential integrability for the stochastic Langevin equation. SIAM Journal on Numerical Analysis , 63(2):1000–1024, 2025

work page 2025
[6]

Cheng, N

X. Cheng, N. S. Chatterji, Y. Abbasi-Yadkori, P. L. Bartlett, and M. I. Jordan. Sharp conver- gence rates for Langevin dynamics in the nonconvex setting.arXiv preprint arXiv:1805.01648, 2018. 28

arXiv 2018

Show all 30 references

[7]

Cheng, N

X. Cheng, N. S. Chatterji, P. L. Bartlett, and M. I. Jordan. Underdamped Langevin MCMC: A non-asymptotic analysis. In Conference on Learning Theory, pages 300–323. PMLR, 2018

2018
[8]

S. Cox, M. Hutzenthaler, and A. Jentzen. Local Lipschitz Continuity in the Initial Value and Strong Completeness for Nonlinear Stochastic Differential Equations , volume 296. American Mathematical Society, 2024

2024
[9]

J. Cui, J. Hong, and D. Sheng. Density function of numerical solution of splitting A VF scheme for stochastic Langevin equation. Mathematics of Computation , 91(337):2283–2333, 2022

2022
[10]

J. Cui, J. Hong, and L. Sun. The stochastic scalar auxiliary variable approach for stochastic nonlinear Klein–Gordon equation. Stochastics and Partial Differential Equations: Analysis and Computations , pages 1–33, 2025

2025
[12]

A. S. Dalalyan and L. Riou-Durand. On sampling from a log-concave density using kinetic Langevin diffusions. Bernoulli, 26(3):1956–1988, 2020

1956
[13]

Eberle, A

A. Eberle, A. Guillin, and R. Zimmer. Couplings and quantitative contraction rates for Langevin dynamics. Annals of Probability, 47(4):1982–2010, 2019

1982
[14]

X. Gao, M. G¨ urb¨ uzbalaban, and L. Zhu. Global convergence of stochastic gradient Hamilto- nian Monte Carlo for nonconvex stochastic optimization: nonasymptotic performance bounds and momentum-based acceleration. Operations Research, 70(5):2931–2947, 2022

2022
[15]

Hutzenthaler and A

M. Hutzenthaler and A. Jentzen. Convergence of the stochastic Euler scheme for locally Lipschitz coefficients. Foundations of Computational Mathematics , 11(6):657–706, 2011

2011
[16]

Hutzenthaler and A

M. Hutzenthaler and A. Jentzen. On a perturbation theory and on strong convergence rates for stochastic ordinary and partial differential equations with nonglobally monotone coefficients. The Annals of Probability , 48(1):53–93, 2020

2020
[17]

Hutzenthaler, A

M. Hutzenthaler, A. Jentzen, and X. Wang. Exponential integrability properties of numer- ical approximation processes for nonlinear stochastic differential equations. Mathematics of Computation, 87(311):1353–1413, 2018

2018
[18]

Johnston, I

T. Johnston, I. Lytras, and S. Sabanis. Kinetic Langevin MCMC sampling without gradient Lipschitz continuity-the strongly convex case. Journal of Complexity , 85:101873, 2024

2024
[19]

Leimkuhler, D

B. Leimkuhler, D. Paulin, and P. A. Whalley. Contraction and convergence rates for dis- cretized kinetic Langevin dynamics. SIAM Journal on Numerical Analysis , 62(3):1226–1258, 2024

2024
[20]

Leimkuhler, A

B. Leimkuhler, A. Sharma, and M. V. Tretyakov. Numerical integrators for confined Langevin dynamics. arXiv preprint arXiv:2404.16584 , 2024

2024
[21]

Leroy, B

A. Leroy, B. Leimkuhler, J. Latz, and D. J. Higham. Adaptive stepsize algorithms for Langevin dynamics. SIAM Journal on Scientific Computing , 46(6):A3574–A3598, 2024. 29

2024
[22]

J. C. Mattingly, A. M. Stuart, and D. J. Higham. Ergodicity for SDEs and approxima- tions: locally Lipschitz vector fields and degenerate noise. Stochastic Processes and Their Applications, 101(2):185–232, 2002

2002
[23]

S. Metzger. A convergent stochastic scalar auxiliary variable method. IMA Journal of Nu- merical Analysis, page drae065, 2024

2024
[24]

S. Metzger. Strong error estimates for a fully discrete SA V scheme for the stochastic Allen–Cahn equation with multiplicative noise. ESAIM: M2AN , 59(4):2305–2327, 2025

2025
[25]

G. A. Pavliotis. Stochastic Processes and Applications: Diffusion Processes, the Fokker– Planck and Langevin Equations . Springer, 2014

2014
[26]

R. Y. Rubinstein and D. P. Kroese. Simulation and the Monte Carlo method . John Wiley & Sons, 2016

2016
[27]

Schuh and P

K. Schuh and P. A. Whalley. Convergence of kinetic Langevin samplers for non-convex potentials. arXiv preprint arXiv:2405.09992 , 2024

2024 arXiv
[28]

Shen and J

J. Shen and J. Xu. Convergence and error analysis for the scalar auxiliary variable (SA V) schemes to gradient flows. SIAM Journal on Numerical Analysis , 56(5):2895–2912, 2018

2018
[29]

J. Shen, J. Xu, and J. Yang. The scalar auxiliary variable (SA V) approach for gradient flows. Journal of Computational Physics , 353:407–416, 2018

2018

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [11]

Dai and X

L. Dai and X. Wang. Perturbation estimates for order-one strong approximations of SDEs without globally monotone coefficients. IMA Journal of Numerical Analysis , page draf034, 2025

work page 2025

[2] [30]

D. Talay. Stochastic Hamiltonian systems: exponential convergence to the invariant measure, and discretization by the implicit Euler scheme.Markov Process. Related Fields, 8(2):163–198, 2002. 30

work page 2002

[3] [1]

Abdulle, G

A. Abdulle, G. Vilmart, and K. C. Zygalakis. Long time accuracy of Lie–Trotter splitting methods for Langevin dynamics. SIAM Journal on Numerical Analysis , 53(1):1–16, 2015

work page 2015

[4] [2]

C.-E. Brehier. Approximation of the invariant distribution for a class of ergodic SDEs with one-sided Lipschitz continuous drift coefficient using an explicit tamed Euler scheme.ESAIM: Probability and Statistics , 27:841–866, 2023

work page 2023

[5] [3]

Y. Cao, J. Lu, and L. Wang. On explicit L2-convergence rate estimate for underdamped Langevin dynamics. Archive for Rational Mechanics and Analysis , 247(5):90, 2023

work page 2023

[6] [4]

S. Cerrai. Second Order PDE’s in Finite and Infinite Dimension: a Probabilistic Approach . Springer, 2001

work page 2001

[7] [5]

C. Chen, T. Dang, J. Hong, and F. Zhang. A new class of splitting methods that preserve ergodicity and exponential integrability for the stochastic Langevin equation. SIAM Journal on Numerical Analysis , 63(2):1000–1024, 2025

work page 2025

[8] [6]

Cheng, N

X. Cheng, N. S. Chatterji, Y. Abbasi-Yadkori, P. L. Bartlett, and M. I. Jordan. Sharp conver- gence rates for Langevin dynamics in the nonconvex setting.arXiv preprint arXiv:1805.01648, 2018. 28

arXiv 2018