betaFeedback

Reading v1View latest version

WorksMathematics

Stochastic Gradient Leimkuhler-Matthews Integrator with Second Order Discretization Error

Abstract

Stochastic gradient Langevin dynamics (SGLD) has a discretization error of first order in the step size hh, even with the full gradient. We study the stochastic gradient Leimkuhler-Matthews (SGLM) integrator, which has the same cost as SGLD, one stochastic gradient per step, and reuses each Gaussian vector in consecutive steps. With the full gradient, SGLM is the Leimkuhler-Matthews (LM) integrator, whose discretization error for the invariant distribution is of second order. For a potential strongly convex outside a ball, we prove the bound O(1Nh+Σ2h2+h4)\mathcal O(\frac{1}{Nh} + Σ^2 h^2 + h^4) on the mean square error (MSE) of time averages of SGLM over NN steps, where ΣΣ bounds the variance of the gradient noise. The proof uses a modified class that absorbs the non-Markovian part of SGLM. The discretization error enters this bound as h4h^4 and the MSE bound of SGLD as h2h^2, while a larger batch reduces only the term Σ2h2Σ^2 h^2 in both bounds. In Bayesian logistic regression on MNIST, SGLM has a smaller estimated MSE than SGLD. The coefficient of hh in the numerical bias of SGLM is given explicitly by the covariance of the gradient noise and the continuous Poisson equation. The coefficient for SGLD contains in addition the Hessian of the potential, which comes from the discretization error. Through Riesz representors, the coefficients of the numerical bias determine the coefficients of the Wasserstein-1 errors of SGLM and SGLD.

Introduction

Sampling from the Boltzmann distribution π(x)∝e−U(x)\pi(x) \propto e^{-U(x)} on Rd\mathbb{R}^{d} is a central task in statistical physics [57, 63] and in Bayesian inference [27, 33, 103]. The simplest stochastic differential equation that leaves π\pi invariant is the overdamped Langevin dynamics

dXt=−∇U(Xt) dt+2 dBt,X0=x∈Rd,(1)\mathrm{d}X_t=-\nabla U(X_t)\,\mathrm{d}t+\sqrt{2}\,\mathrm{d}B_t,\qquad X_0=x\in\mathbb{R}^{d}, \tag*{(1)}

where (Bt)t≥0(B_t)_{t\ge0} is a standard Brownian motion in Rd\mathbb{R}^{d} and UU a confining potential; time averages along the solution (Xt)t≥0(X_t)_{t\ge0} approximate expectations under π\pi in the long-time limit [53, 90]. A numerical integrator of (1) with step size h>0h>0 produces samples (Zk)k≥0(Z_k)_{k\ge0} that approximate the invariant distribution π\pi; such integrators form the basis of Langevin Monte Carlo [27, 33, 34, 73, 100]. Characterizing the accuracy of these samples is a basic question of numerical analysis.

The mean square error (MSE) of time averages is a natural criterion for this accuracy. For a test function ff, with π(f):=∫Rdf dπ\pi(f):=\int_{\mathbb{R}^{d}}f\,\mathrm{d}\pi, the MSE of the time average over NN steps is

MSE⁡(N,h):=E[∣1N∑k=0N−1f(Zk)−π(f)∣2].(2)\operatorname{MSE}(N,h):=\mathbb{E}\left[\left|\frac{1}{N}\sum_{k=0}^{N-1}f(Z_k)-\pi(f)\right|^{2}\right]. \tag*{(2)}

The Wasserstein distance, the Kullback–Leibler (KL) divergence and the total variation distance compare the law of the samples with π\pi, and are hard to evaluate in high dimension. The MSE instead concerns a single numerical trajectory, and accounts for both the fluctuation of the time average and the numerical bias, the deviation of its long-time limit from π(f)\pi(f) [18, 77, 99, 102].

In Bayesian inference on large data sets, ∇U\nabla U is replaced by an unbiased stochastic estimator ∇U^(x;ξ)\widehat{\nabla U}(x;\xi) computed from a mini-batch, with ξ\xi a random variable independent of the Brownian motion [18, 99, 102, 103]. The standard sampler of this kind is stochastic gradient Langevin dynamics (SGLD) [103], the Euler–Maruyama (EM) integrator [50, 80] of (1) with the stochastic gradient,

Zk+1=Zk−h∇U^(Zk;ξk)+2h χk+1.Z_{k+1}=Z_k-h\widehat{\nabla U}(Z_k;\xi_k)+\sqrt{2h}\,\chi_{k+1}.

where (χk)k≥0(\chi_k)_{k\geq0} is a sequence of independent standard Gaussian vectors in Rd\mathbb{R}^d and ξk\xi_k is drawn afresh at every step. The discretization error of (SGLD) is of first order in hh: its numerical bias is of order hh even with the full gradient (§ 2.2). The existing stochastic gradient integrators that preserve a second order discretization error are built on Hamiltonian Monte Carlo [18, 22, 72] or on the underdamped Langevin dynamics [44], whose splitting integrators are of second order with one gradient evaluation per step [81]; the underdamped dynamics with stochastic gradients is studied in [58, 71, 76, 88].

We study the stochastic gradient Leimkuhler–Matthews (SGLM) integrator, which has the cost of (SGLD), one stochastic gradient per step, and differs from it in only one respect: the Gaussian vector χk+1\chi_{k+1} is replaced by the average of χk\chi_k and χk+1\chi_{k+1}, so that each Gaussian vector is used twice, in consecutive steps,

Zk+1=Zk−h∇U^(Zk;ξk)+h2(χk+χk+1).Z_{k+1}=Z_k-h\widehat{\nabla U}(Z_k;\xi_k)+\sqrt{\frac{h}{2}}(\chi_k+\chi_{k+1}).

With the full gradient, SGLM is the Leimkuhler–Matthews (LM) integrator [56, 101],

Zk+1=Zk−h∇U(Zk)+h2(χk+χk+1).Z_{k+1}=Z_k-h\nabla U(Z_k)+\sqrt{\frac{h}{2}}(\chi_k+\chi_{k+1}).

The reuse of the Gaussian vector gives LM a second order discretization error for the invariant distribution π\pi [56, 59, 101], although its samples (Zk)k≥0(Z_k)_{k\geq0} do not form a Markov chain and its weak error on finite time intervals is of first order. LM evaluates ∇U\nabla U once per step, while almost all other integrators of (1) with a second order discretization error are of predictor–corrector type, with at least two sequential evaluations of ∇U\nabla U per step (§ 2.1). With a stochastic gradient, each of these evaluations carries its own noise. This raises the question whether SGLM preserves the second order discretization error of LM.

We provide a positive answer to this question. Our contributions are summarized as follows:

  1. SGLM separates the stochastic gradient error from the discretization error.

Under regularity conditions on UU, ff, and the stochastic gradient that require strong convexity of UU only outside a ball, we prove in Theorem 3 the MSE bound for SGLM,

MSE(N,h)≤C(1Nh+Σ2h2+h4).(3)\text{MSE}(N,h)\leq C\left(\frac{1}{Nh}+\Sigma^2h^2+h^4\right). \tag*{(3)}

where Σ\Sigma bounds the variance of the stochastic gradient, and CC depends on the regularity constants of UU and ff, the dimension dd and the initial value Z0Z_0, but not on NN, hh or Σ\Sigma. In (3), 1Nh\frac{1}{Nh} comes from the fluctuation of the time average, Σ2h2\Sigma^2h^2 from the square of the stochastic gradient error, and h4h^4 from the square of the discretization error. The same separation holds for the UBU integrator with stochastic gradients (SG-UBU) [71], and was previously known only for integrators of the underdamped dynamics [71, 87]. The (SGLD) integrator does not have it, since its stochastic gradient error and its discretization error are both of first order in hh. With the full gradient, Σ=0\Sigma=0 and (3) yields the MSE bound 1Nh+h4\frac{1}{Nh}+h^4 of LM. To our knowledge, this is also the first MSE bound for the LM integrator.

The separation in (3) is advantageous in the mini-batch regime: if each step averages PP independent stochastic gradients, the variance bound Σ\Sigma is replaced by Σ/P\Sigma/P, so the contribution of the gradient noise to the MSE becomes Σ2h2/P2\Sigma^2h^2/P^2, with CC in (3) independent of PP. Increasing the batch size PP at a fixed step size hh therefore reduces the stochastic gradient error. By contrast, the MSE of (SGLD) keeps the term h2h^{2} from the square of its discretization error at every batch size, and only a smaller hh reduces it. Since the gradients of a batch are evaluated in parallel while the steps are sequential, SGLM with a large batch size PP has a smaller error than SGLD under the same cost.

The proof of (3) runs on the Markovian form of (SGLM), a Markov chain (Yk)k≥0(Y_k)_{k\geq0} with Zk=Yk+h/2 χkZ_k=Y_k+\sqrt{h/2}\,\chi_k. It builds on the discrete Poisson equation of our earlier work [71, 106], combined with a modified class of test function, potential, generator and discrete Poisson solution that absorbs the non-Markovian part of the integrator, the Gaussian shift h/2 χk\sqrt{h/2}\,\chi_k.

The coefficients of the numerical bias are explicit.

The time averages of ff along (SGLM) converge almost surely to πh(f)\pi_h(f), for a numerical invariant distribution πh\pi_h independent of the initial value, and the numerical bias πh(f)−π(f)\pi_h(f)-\pi(f) is of order hh with an explicit coefficient given in Corollary 2,

(SGLM, numerical bias)lim⁡h→0πh(f)−π(f)h=12π(∇2ϕ0(x):C(x)),(4)\text{(SGLM, numerical bias)}\qquad\lim_{h\to0}\frac{\pi_h(f)-\pi(f)}{h}=\frac{1}{2}\pi\left(\nabla^{2}\phi_0(x):C(x)\right), \tag*{(4)}

where C(x)C(x) is the covariance of the gradient noise ∇U^(x;ξ)−∇U(x)\nabla\widehat{U}(x;\xi)-\nabla U(x), ϕ0\phi_0 solves the continuous Poisson equation (14), −Lϕ0=f−π(f)-\mathcal{L}\phi_0=f-\pi(f) with L\mathcal{L} the generator of (1), and :: is the full contraction of two matrices. The coefficient comes entirely from the gradient noise: with the full gradient, C(x)=0C(x)=0 and the numerical bias of LM has no linear part. The gradient noise enters only through the local weak error of a single step, through its covariance C(x)C(x), as for SG-UBU in [71 Section 3.4].

The discrete Poisson equation also characterizes the numerical bias of (SGLD). Its time averages converge to πˉh(f)\bar{\pi}_h(f), with πˉh\bar{\pi}_h the invariant distribution of SGLD, and by Corollary 3

(SGLD, numerical bias)lim⁡h→0πˉh(f)−π(f)h=12π(∇2ϕ0(x):(∇2U(x)+C(x))),(5)\text{(SGLD, numerical bias)}\qquad\lim_{h\to0}\frac{\bar{\pi}_h(f)-\pi(f)}{h}=\frac{1}{2}\pi\left(\nabla^{2}\phi_0(x):\left(\nabla^{2}U(x)+C(x)\right)\right), \tag*{(5)}

a coefficient studied formally for d=1d=1 in [102]. Besides the covariance C(x)C(x), the coefficient of SGLD carries the Hessian ∇2U(x)\nabla^{2}U(x), which comes from the discretization error of EM. With the full gradient, C(x)=0C(x)=0 and (5) gives the coefficient of EM; with a batch of PP independent stochastic gradients, the coefficient of SGLM is divided by PP, while that of SGLD tends to the coefficient of EM as PP grows. In (4) and (5), the gradient noise and the discretization error both enter linearly through ∇2ϕ0\nabla^{2}\phi_0, so the solution ϕ0\phi_0 of the continuous Poisson equation is an essential structure of the numerical bias of the two stochastic gradient integrators, not only a tool for upper bounds on the MSE.

The coefficients of the Wasserstein-1 error are explicit.

The suprema of the absolute values of the coefficients in (4) and (5) over the test functions with ∥∇f∥∞≤1\lVert\nabla f\rVert_{\infty}\leq1 are the limits of the normalized Wasserstein-1 errors W1(πh,π)/h\mathcal{W}_1(\pi_h,\pi)/h and W1(πˉh,π)/h\mathcal{W}_1(\bar{\pi}_h,\pi)/h as h→0h\to0 (Corollary 4):

(SGLM, W1 coefficient)lim⁡h→0W1(πh,π)h=12∥ψCπ∥KR,(6)\text{(SGLM, }\mathcal{W}_1\text{ coefficient)}\qquad\lim_{h\to0}\frac{\mathcal{W}_1(\pi_h,\pi)}{h}=\frac{1}{2}\lVert\psi_C\pi\rVert_{\mathrm{KR}}, \tag*{(6)}

References

  1. [1]Assyr Abdulle, Gilles Vilmart, and Konstantinos C. Zygalakis. High order numerical approximation of the invariant measure of ergodic SDEs. SIAM Journal on Numerical Analysis, 52(4):1600–1622, 2014. doi: 10.1137/130935616.DOI
  2. [2]Ömer Deniz Akyildiz and Sotirios Sabanis. Nonasymptotic analysis of stochastic gradient Hamiltonian Monte Carlo under local conditions for nonconvex optimization. Journal of Machine Learning Research, 25(113):1–34, 2024.
  3. [3]Jason M. Altschuler and Sinho Chewi. Faster high-accuracy log-concave sampling via algorithmic warm starts. Journal of the ACM, 71(3):1–55, 2024. doi: 10.1145/3653446.DOI
  4. [4]Jack Baker, Paul Fearnhead, Emily B. Fox, and Christopher Nemeth. Control variates for stochastic gradient MCMC. Statistics and Computing, 29(3):599–615, 2019. doi: 10.1007/s11222-018-9826-2.DOI
  5. [5]Dominique Bakry, Franck Barthe, Patrick Cattiaux, and Arnaud Guillin. A simple proof of the Poincaré inequality for a large class of probability measures including the log-concave case. Electronic Communications in Probability, 13:60–66, 2008. doi: 10.1214/ECP.v13-1352.DOI
  6. [6]Dominique Bakry, Ivan Gentil, and Michel Ledoux. Analysis and Geometry of Markov Diffusion Operators, volume 348 of Grundlehren der mathematischen Wissenschaften. Springer, 2014. doi: 10.1007/978-3-319-00227-9.DOI
  7. [7]Mathias Barkhagen, Ngoc Huy Chau, Éric Moulines, Miklós Rásonyi, Sotirios Sabanis, and Ying Zhang. On stochastic gradient Langevin dynamics with dependent data streams in the logconcave case. Bernoulli, 27(1):1–33, 2021. doi: 10.3150/19-BEJ1187.DOI
  8. [8]Eli Bingham, Jonathan P. Chen, Martin Jankowiak, Fritz Obermeyer, Neeraj Pradhan, Theofanis Karaletsos, Rohit Singh, Paul Szerlip, Paul Horsfall, and Noah D. Goodman. Pyro: Deep universal probabilistic programming. Journal of Machine Learning Research, 20(28):1–6, 2019.arxiv.org/abs/1810.09538
  9. [9]Nicolas Brosse, Alain Durmus, and Éric Moulines. The promises and pitfalls of stochastic gradient Langevin dynamics. In Advances in Neural Information Processing Systems, volume 31, 2018.arxiv.org/abs/1811.10072
  10. [10]Nicolas Brosse, Alain Durmus, Éric Moulines, and Sotirios Sabanis. The tamed unadjusted Langevin algorithm. Stochastic Processes and their Applications, 129(10):3638–3663, 2019. doi: 10.1016/j.spa.2018.10.002.
  11. [11]Giovanni Bussi and Michele Parrinello. Accurate sampling using Langevin dynamics. Physical Review E, 75(5):056707, 2007. doi: 10.1103/PhysRevE.75.056707.DOI
  12. [12]Yu Cao and Eric Vanden-Eijnden. Learning optimal flows for non-equilibrium importance sampling. In Advances in Neural Information Processing Systems, volume 35, pages 12352–12364, 2022.arxiv.org/abs/2206.09908
  13. [13]Neil K. Chada, Benedict Leimkuhler, Daniel Paulin, and Peter A. Whalley. Unbiased kinetic Langevin Monte Carlo with inexact gradients. The Annals of Statistics, 54(4):1895–1922, 2026. doi: 10.1214/26-AOS2630.DOI
  14. [14]Martin Chak and Pierre Monmarché. Reflection coupling for unadjusted generalized Hamiltonian Monte Carlo in the nonconvex stochastic gradient case. IMA Journal of Numerical Analysis, 46(4):2216–2266, 2026. doi: 10.1093/imanum/draf045.DOI
  15. [15]Niladri S. Chatterji, Nicolas Flammarion, Yi-An Ma, Peter L. Bartlett, and Michael I. Jordan. On the theory of variance reduction for stochastic gradient Monte Carlo. In International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 764–773, 2018.arxiv.org/abs/1802.05431
  16. [16]Niladri S. Chatterji, Peter L. Bartlett, and Philip M. Long. Oracle lower bounds for stochastic gradient sampling algorithms. Bernoulli, 28(2):1074–1092, 2022. doi: 10.3150/21-BEJ1377.DOI
  17. [17]Ngoc Huy Chau, Éric Moulines, Miklós Rásonyi, Sotirios Sabanis, and Ying Zhang. On stochastic gradient Langevin dynamics with dependent data streams: the fully nonconvex case. SIAM Journal on Mathematics of Data Science, 3(3):959–986, 2021. doi: 10.1137/20M1355392.DOI
  18. [18]Changyou Chen, Nan Ding, and Lawrence Carin. On the convergence of stochastic gradient MCMC algorithms with high-order integrators. In Advances in Neural Information Processing Systems, volume 28, 2015.
  19. [19]Changyou Chen, Wenlin Wang, Yizhe Zhang, Qinliang Su, and Lawrence Carin. A convergence analysis for a class of practical variance-reduction stochastic gradient MCMC. Science China Information Sciences, 62(1):12101, 2019. doi: 10.1007/s11432-018-9656-y.DOI
  20. [20]Fan Chen, Sinho Chewi, Constantinos Daskalakis, and Alexander Rakhlin. High-accuracy sampling for diffusion models and log-concave distributions. In International Conference on Machine Learning, volume 306 of Proceedings of Machine Learning Research, pages 13835–13869, 2026.
  21. [21]Fan Chen, Sinho Chewi, Constantinos Daskalakis, and Alexander Rakhlin. High-accuracy log-concave sampling with stochastic queries. In Conference on Learning Theory, volume 336 of Proceedings of Machine Learning Research, pages 1341–1372, 2026.
  22. [22]Tianqi Chen, Emily B. Fox, and Carlos Guestrin. Stochastic gradient Hamiltonian Monte Carlo. In International Conference on Machine Learning, volume 32 of Proceedings of Machine Learning Research, pages 1683–1691, 2014.
  23. [23]Yifan Chen, Xiaoou Cheng, Jonathan Niles-Weed, and Jonathan Weare. Convergence of unadjusted Langevin in high dimensions: delocalization of bias. Communications on Pure and Applied Mathematics, 79(6):1467–1491, 2026. doi: 10.1002/cpa.70032.DOI
  24. [24]Yifan Chen, Xiaoou Cheng, Jonathan Niles-Weed, and Jonathan Weare. Delocalization of bias in unadjusted Hamiltonian Monte Carlo and underdamped Langevin. arXiv:2607.15208, 2026.arxiv.org/abs/2607.15208
  25. [25]Yongxin Chen, Sinho Chewi, Adil Salim, and Andre Wibisono. Improved analysis for a proximal algorithm for sampling. In Conference on Learning Theory, volume 178 of Proceedings of Machine Learning Research, pages 2984–3014, 2022.arxiv.org/abs/2202.06386
  26. [27]Arnak S. Dalalyan and Avetik Karagulyan. User-friendly guarantees for the Langevin Monte Carlo with inaccurate gradient. Stochastic Processes and their Applications, 129(12):5278–5311, 2019. doi: 10.1016/j.spa.2019.02.016.DOI
  27. [28]Aniket Das, Dheeraj M. Nagaraj, and Anant Raj. Utilising the CLT structure in stochastic gradient based sampling: improved analysis and faster algorithms. In Conference on Learning Theory, volume 195 of Proceedings of Machine Learning Research, pages 4072–4129, 2023.arxiv.org/abs/2206.03792
  28. [29]Kristian Debrabant. Runge–Kutta methods for third order weak approximation of SDEs with multidimensional additive noise. BIT Numerical Mathematics, 50(3):541–558, 2010. doi: 10.1007/s10543-010-0276-2.DOI
  29. [30]Arnaud Debussche and Erwan Faou. Weak backward error analysis for SDEs. SIAM Journal on Numerical Analysis, 50(3):1735–1752, 2012. doi: 10.1137/110831544.DOI
  30. [31]Nan Ding, Youhan Fang, Ryan Babbush, Changyou Chen, Robert D. Skeel, and Hartmut Neven. Bayesian sampling using stochastic gradient thermostats. In Advances in Neural Information Processing Systems, volume 27, 2014.
  31. [32]Kumar Avinava Dubey, Sashank J. Reddi, Sinead A. Williamson, Barnabás Póczos, Alexander J. Smola, and Eric P. Xing. Variance reduction in stochastic gradient Langevin dynamics. In Advances in Neural Information Processing Systems, volume 29, 2016.pubmed.ncbi.nlm.nih.gov/28713210
  32. [33]Alain Durmus and Éric Moulines. Nonasymptotic convergence analysis for the unadjusted Langevin algorithm. The Annals of Applied Probability, 27(3):1551–1587, 2017. doi: 10.1214/16-AAP1238.DOI
  33. [34]Alain Durmus and Éric Moulines. High-dimensional Bayesian inference via the unadjusted Langevin algorithm. Bernoulli, 25(4A):2854–2882, 2019. doi: 10.3150/18-BEJ1073.DOI
  34. [35]Alain Durmus, Umut Simsekli, Éric Moulines, Roland Badeau, and Gaël Richard. Stochastic gradient Richardson–Romberg Markov chain Monte Carlo. In Advances in Neural Information Processing Systems, volume 29, 2016.
  35. [36]Alain Durmus, Szymon Majewski, and Błażej Miasojedow. Analysis of Langevin Monte Carlo via convex optimization. Journal of Machine Learning Research, 20(73):1–46, 2019.
  36. [37]Andreas Eberle. Reflection couplings and contraction rates for diffusions. Probability Theory and Related Fields, 166(3-4):851–886, 2016. doi: 10.1007/s00440-015-0673-1.DOI
  37. [38]K. David Elworthy and Xue-Mei Li. Formulae for the derivatives of heat semigroups. Journal of Functional Analysis, 125(1):252–286, 1994. doi: 10.1006/jfan.1994.1124.DOI
  38. [39]Thomas O. Gallouët, Andrea Natale, and Gabriele Todeschi. From geodesic extrapolation to a variational BDF2 scheme for Wasserstein gradient flows. Mathematics of Computation, 93(350):2769–2810, 2024. doi: 10.1090/mcom/3951.DOI
  39. [40]Thomas O. Gallouët, Andrea Natale, and Gabriele Todeschi. Metric extrapolation in the Wasserstein space. Calculus of Variations and Partial Differential Equations, 64(5):147, 2025. doi: 10.1007/s00526-025-02996-6.DOI
  40. [41]Xuefeng Gao, Mert Gürbüzbalaban, and Lingjing Zhu. Global convergence of stochastic gradient Hamiltonian Monte Carlo for nonconvex stochastic optimization: nonasymptotic performance bounds and momentum-based acceleration. Operations Research, 70(5):2931–2947, 2022. doi: 10.1287/opre.2021.2162.DOI
  41. [42]Michael B. Giles, Tigran Nagapetyan, Lukasz Szpruch, Sebastian J. Vollmer, and Konstantinos C. Zygalakis. Multilevel Monte Carlo for scalable Bayesian computations. arXiv:1609.06144, 2016.arxiv.org/abs/1609.06144
  42. [43]Michael B. Giles, Mateusz B. Majka, Lukasz Szpruch, Sebastian J. Vollmer, and Konstantinos C. Zygalakis. Multi-level Monte Carlo methods for the approximation of invariant measures of stochastic differential equations. Statistics and Computing, 30(3):507–524, 2020. doi: 10.1007/s11222-019-09890-0.DOI
  43. [44]Nicolai Gouraud, Pierre Le Bris, Adrien Majka, and Pierre Monmarché. HMC and underdamped Langevin united in the unadjusted convex smooth case. SIAM/ASA Journal on Uncertainty Quantification, 13(1):278–303, 2025. doi: 10.1137/23M1608963.DOI
  44. [45]Melville S. Green. Markoff random processes and the statistical mechanics of time-dependent phenomena. II. Irreversible processes in fluids. The Journal of Chemical Physics, 22(3):398–413, 1954. doi: 10.1063/1.1740082.DOI
  45. [46]Martin Hairer and Andrew J. Majda. A simple framework to justify linear response theory. Nonlinearity, 23(4):909–922, 2010. doi: 10.1088/0951-7715/23/4/008.DOI
  46. [47]Martin Hairer and Jonathan C. Mattingly. Yet another look at Harris’ ergodic theorem for Markov chains. In Robert C. Dalang, Marco Dozzi, and Francesco Russo, editors, Seminar on Stochastic Analysis, Random Fields and Applications VI, volume 63 of Progress in Probability, pages 109–117. Birkhäuser/Springer Basel, 2011. doi: 10.1007/978-3-0348-0021-1_7.DOI
  47. [48]Matthew D. Hoffman and Andrew Gelman. The No-U-Turn sampler: adaptively setting path lengths in Hamiltonian Monte Carlo. Journal of Machine Learning Research, 15(47):1593–1623, 2014.DOI
  48. [49]Yuri Kinoshita and Taiji Suzuki. Improved convergence rate of stochastic gradient Langevin dynamics with variance reduction and its application to optimization. In Advances in Neural Information Processing Systems, volume 35, 2022.
  49. [50]Peter E. Kloeden and Eckhard Platen. Numerical Solution of Stochastic Differential Equations, volume 23 of Applications of Mathematics (New York). Springer, 1992. doi: 10.1007/978-3-662-12616-5.DOI
  50. [51]Marie Kopec. Weak backward error analysis for overdamped Langevin processes. IMA Journal of Numerical Analysis, 35(2):583–614, 2015. doi: 10.1093/imanum/dru016.DOI
  51. [52]Ryogo Kubo. Statistical-mechanical theory of irreversible processes. I. General theory and simple applications to magnetic and conduction problems. Journal of the Physical Society of Japan, 12(6):570–586, 1957. doi: 10.1143/JPSJ.12.570.DOI
  52. [53]Damien Lamberton and Gilles Pagès. Recursive computation of the invariant distribution of a diffusion. Bernoulli, 8(3):367–405, 2002.
  53. [54]Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998. doi: 10.1109/5.726791.DOI
  54. [55]Yin Tat Lee, Ruoqi Shen, and Kevin Tian. Structured logconcave sampling with a restricted Gaussian oracle. In Conference on Learning Theory, volume 134 of Proceedings of Machine Learning Research, pages 2993–3050, 2021.arxiv.org/abs/2010.03106
  55. [56]Benedict Leimkuhler and Charles Matthews. Rational construction of stochastic numerical methods for molecular sampling. Applied Mathematics Research eXpress, 2013(1):34–56, 2013. doi: 10.1093/amrx/abs010.DOI
  56. [57]Benedict Leimkuhler and Charles Matthews. Molecular Dynamics: With Deterministic and Stochastic Numerical Methods, volume 39 of Interdisciplinary Applied Mathematics. Springer, 2015. doi: 10.1007/978-3-319-16375-8.DOI
  57. [58]Benedict Leimkuhler and Xiaocheng Shang. Adaptive thermostats for noisy gradient systems. SIAM Journal on Scientific Computing, 38(2):A712–A736, 2016. doi: 10.1137/15M102318X.DOI
  58. [59]Benedict Leimkuhler, Charles Matthews, and Michael V. Tretyakov. On the long-time integration of stochastic gradient systems. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 470(2170):20140120, 2014. doi: 10.1098/rspa.2014.0120.DOI
  59. [60]Benedict Leimkuhler, Charles Matthews, and Gabriel Stoltz. The computation of averages from equilibrium and nonequilibrium Langevin molecular dynamics. IMA Journal of Numerical Analysis, 36(1):13–79, 2016. doi: 10.1093/imanum/dru056.DOI
  60. [61]Benedict Leimkuhler, Matthias Sachs, and Gabriel Stoltz. Hypocoercivity properties of adaptive Langevin dynamics. SIAM Journal on Applied Mathematics, 80(3):1197–1222, 2020. doi: 10.1137/19M1291649.DOI
  61. [62]Benedict Leimkuhler, Daniel Paulin, and Peter A. Whalley. Contraction rate estimates of stochastic gradient kinetic Langevin integrators. ESAIM: Mathematical Modelling and Numerical Analysis, 58(6):2255–2286, 2024. doi: 10.1051/m2an/2024038.DOI
  62. [63]Tony Lelièvre and Gabriel Stoltz. Partial differential equations and stochastic methods in molecular dynamics. Acta Numerica, 25:681–880, 2016. doi: 10.1017/S0962492916000039.DOI
  63. [64]Vincent Lemaire, Gilles Pagès, and Fabien Panloup. Invariant measure of duplicated diffusions and application to Richardson–Romberg extrapolation. Annales de l’Institut Henri Poincaré, Probabilités et Statistiques, 51(4):1562–1596, 2015. doi: 10.1214/13-AIHP591.DOI
  64. [65]Lei Li and Yuliang Wang. A sharp uniform-in-time error estimate for stochastic gradient Langevin dynamics. CSIAM Transactions on Applied Mathematics, 6(4):711–759, 2025. doi: 10.4208/csiam-am.SO-2024-0039.
  65. [66]Lei Li, Jian-Guo Liu, and Yuliang Wang. Geometric ergodicity of SGLD via reflection coupling. Stochastics and Dynamics, 24(5):2450035, 2024. doi: 10.1142/S0219493724500357.DOI
  66. [67]Ruilin Li, Xin Wang, Hongyuan Zha, and Molei Tao. Improving sampling accuracy of stochastic gradient MCMC methods via non-uniform subsampling of gradients. Discrete and Continuous Dynamical Systems – Series S, 16(2):329–360, 2023. doi: 10.3934/dcdss.2021157.
  67. [68]Wuchen Li, Ernest K. Ryu, Stanley Osher, Wotao Yin, and Wilfrid Gangbo. A parallel method for Earth Mover’s distance. Journal of Scientific Computing, 75(1):182–197, 2018. doi: 10.1007/s10915-017-0529-1.DOI
  68. [69]Xuechen Li, Yi Wu, Lester Mackey, and Murat A. Erdőgdu. Stochastic Runge–Kutta accelerates Langevin Monte Carlo and beyond. In Advances in Neural Information Processing Systems, volume 32, 2019.
  69. [70]Zhize Li, Tianyi Zhang, Shuyu Cheng, Jun Zhu, and Jian Li. Stochastic gradient Hamiltonian Monte Carlo with variance reduction for Bayesian inference. Machine Learning, 108(8–9):1701–1727, 2019. doi: 10.1007/s10994-019-05825-y.DOI
  70. [71]Jianfeng Lu, Xuda Ye, and Zhennan Zhou. Mean square error analysis of stochastic gradient and variance-reduced sampling algorithms. arXiv:2511.04413, 2025.arxiv.org/abs/2511.04413
  71. [72]Yi-An Ma, Tiangi Chen, and Emily B. Fox. A complete recipe for stochastic gradient MCMC. In Advances in Neural Information Processing Systems, volume 28, 2015.
  72. [73]Mateusz B. Majka, Aleksandar Mijatović, and Lukasz Szpruch. Nonasymptotic bounds for sampling algorithms without log-concavity. The Annals of Applied Probability, 30(4):1534–1581, 2020. doi: 10.1214/19-AAP1535.DOI
  73. [74]Mateusz B. Majka, Tigran Nagapetyan, Lukasz Szpruch, Yue Wu, and Danqi Zhuang. On the computational cost of stochastic gradient Langevin dynamics. arXiv:2609.17750, 2026.arxiv.org/abs/2609.17750
  74. [75]Riccardo Mannella. Numerical integration of stochastic differential equations: the Heun algorithm revisited and the Itô–Stratonovich calculus. Entropy, 27(9):910, 2025. doi: 10.3390/e27090910.DOI
  75. [76]Charles Matthews and Jonathan Weare. Langevin Markov chain Monte Carlo with stochastic gradients. arXiv:1805.08863, 2018.arxiv.org/abs/1805.08863
  76. [77]Jonathan C. Mattingly, Andrew M. Stuart, and Michael V. Tretyakov. Convergence of numerical time-averaging and stationary measures via Poisson equations. SIAM Journal on Numerical Analysis, 48(2):552–577, 2010. doi: 10.1137/090770527.DOI
  77. [78]Sean P. Meyn and Richard L. Tweedie. Stability of Markovian processes I: Criteria for discrete-time chains. Advances in Applied Probability, 24(3):542–574, 1992. doi: 10.2307/1427479.DOI
  78. [79]Sean P. Meyn and Richard L. Tweedie. Markov Chains and Stochastic Stability. Springer-Verlag, London, 1993. doi: 10.1007/978-1-4471-3267-7.DOI
  79. [80]Grigori N. Milstein and Michael V. Tretyakov. Stochastic Numerics for Mathematical Physics. Scientific Computation. Springer, 2004. doi: 10.1007/978-3-662-10063-9.DOI
  80. [81]Pierre Monmarché. High-dimensional MCMC with a standard splitting scheme for the underdamped Langevin diffusion. Electronic Journal of Statistics, 15(2):4117–4166, 2021. doi: 10.1214/21-EJS1888.DOI
  81. [82]Tigran Nagapetyan, Andrew B. Duncan, Leonard Hasenclever, Sebastian J. Vollmer, Lukasz Szpruch, and Konstantinos C. Zygalakis. The true cost of stochastic gradient Langevin dynamics. arXiv:1706.02692, 2017.arxiv.org/abs/1706.02692
  82. [83]Kirill Neklyudov, Rob Brekelmans, Daniel Severo, and Alireza Makhzani. Action matching: Learning stochastic dynamics from samples. In International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 25858–25889, 2023.arxiv.org/abs/2210.06662
  83. [84]Christopher Nemeth and Paul Fearnhead. Stochastic gradient Markov chain Monte Carlo. Journal of the American Statistical Association, 116(533):433–450, 2021. doi: 10.1080/01621459.2020.1847120.DOI
  84. [85]Daniel Paulin and Peter A. Whalley. Correction to “Wasserstein distance estimates for the distributions of numerical approximations to ergodic stochastic differential equations”. Journal of Machine Learning Research, 25(376):1–9, 2024.arxiv.org/abs/2402.08711
  85. [86]Daniel Paulin and Peter A. Whalley. Theoretical guarantees for stochastic gradient Langevin dynamics. arXiv:2610.01651, 2026.arxiv.org/abs/2610.01651
  86. [87]Daniel Paulin and Peter A. Whalley. Theoretical guarantees for stochastic gradient sampling methods via Gaussian convolution inequalities. arXiv:2604.24632, 2026.arxiv.org/abs/2604.24632
  87. [88]Daniel Paulin, Peter A. Whalley, Neil K. Chada, and Benedict Leimkuhler. Sampling from Bayesian neural network posteriors with symmetric minibatch splitting Langevin dynamics. In International Conference on Artificial Intelligence and Statistics, volume 258 of Proceedings of Machine Learning Research, pages 5014–5022, 2025.arxiv.org/abs/2410.19780
  88. [89]Maxim Raginsky, Alexander Rakhlin, and Matus Telgarsky. Non-convex learning via stochastic gradient Langevin dynamics: a nonasymptotic analysis. In Conference on Learning Theory, volume 65 of Proceedings of Machine Learning Research, pages 1674–1703, 2017.arxiv.org/abs/1702.03849
  89. [90]Gareth O. Roberts and Richard L. Tweedie. Exponential convergence of Langevin distributions and their discrete approximations. Bernoulli, 2(4):341–363, 1996. doi: 10.2307/3318418.DOI
  90. [91]Andreas Rößler. Runge–Kutta methods for the strong approximation of solutions of stochastic differential equations. SIAM Journal on Numerical Analysis, 48(3):922–952, 2010. doi: 10.1137/09076636X.
  91. [92]Jesús María Sanz-Serna and Konstantinos C. Zygalakis. Wasserstein distance estimates for the distributions of numerical approximations to ergodic stochastic differential equations. Journal of Machine Learning Research, 22(242):1–37, 2021.
  92. [93]Issei Sato and Hiroshi Nakagawa. Approximation analysis of stochastic gradient Langevin dynamics by using Fokker–Planck equation and Ito process. In International Conference on Machine Learning, volume 32 of Proceedings of Machine Learning Research, pages 982–990, 2014.
  93. [94]Inass Sekkat and Gabriel Stoltz. Mini-batching error and adaptive Langevin dynamics. Journal of Machine Learning Research, 24(329):1–58, 2023.
  94. [95]Xiaocheng Shang, Zhanxing Zhu, Benedict Leimkuhler, and Amos J. Storkey. Covariance-controlled adaptive Langevin thermostat for large-scale Bayesian sampling. In Advances in Neural Information Processing Systems, volume 28, 2015.arxiv.org/abs/1510.08692
  95. [96]Luke Shaw and Peter A. Whalley. Random reshuffling for stochastic gradient Langevin dynamics. arXiv:2501.16055, 2025.arxiv.org/abs/2501.16055
  96. [97]Denis Talay. Second-order discretization schemes of stochastic differential systems for the computation of the invariant law. Stochastics and Stochastic Reports, 29(1):13–36, 1990. doi: 10.1080/17442509008833606.DOI
  97. [98]Denis Talay and Luciano Tubaro. Expansion of the global error for numerical schemes solving stochastic differential equations. Stochastic Analysis and Applications, 8(4):483–509, 1990. doi: 10.1080/07362999008809220.DOI
  98. [99]Yee Whye Teh, Alexandre H. Thiery, and Sebastian J. Vollmer. Consistency and fluctuations for stochastic gradient Langevin dynamics. Journal of Machine Learning Research, 17(7):1–33, 2016.arxiv.org/abs/1409.0578
  99. [100]Santosh S. Vempala and Andre Wibisono. Rapid convergence of the unadjusted Langevin algorithm: isoperimetry suffices. In Advances in Neural Information Processing Systems, volume 32, 2019.arxiv.org/abs/1903.08568
  100. [101]Gilles Vilmart. Postprocessed integrators for the high order integration of ergodic SDEs. SIAM Journal on Scientific Computing, 37(1):A201–A220, 2015. doi: 10.1137/140974328.DOI
  101. [102]Sebastian J. Vollmer, Konstantinos C. Zygalakis, and Yee Whye Teh. Exploration of the (non-)asymptotic bias and variance of stochastic gradient Langevin dynamics. Journal of Machine Learning Research, 17(159):1–48, 2016.DOI
  102. [103]Max Welling and Yee Whye Teh. Bayesian learning via stochastic gradient Langevin dynamics. In International Conference on Machine Learning, pages 681–688, 2011.
  103. [104]Pan Xu, Jinghui Chen, Difan Zou, and Quanquan Gu. Global convergence of Langevin dynamics based algorithms for nonconvex optimization. In Advances in Neural Information Processing Systems, volume 31, 2018.arxiv.org/abs/1707.06618
  104. [105]Bin Yang and Xiaojie Wang. Accelerating Langevin Monte Carlo via efficient stochastic Runge–Kutta methods beyond log-concavity. In International Conference on Machine Learning, volume 306 of Proceedings of Machine Learning Research, pages 146170–146197, 2026.
  105. [106]Xuda Ye. Mean square error analysis of stochastic Runge–Kutta integrators. arXiv:2609.11528, 2026.arxiv.org/abs/2609.11528
  106. [107]Ying Zhang, Ömer Deniz Akyildiz, Theodorus Damoulas, and Sotirios Sabanis. Nonasymptotic estimates for stochastic gradient Langevin dynamics under local conditions in nonconvex optimization. Applied Mathematics and Optimization, 87(2):25, 2023. doi: 10.1007/s00245-022-09932-6.DOI
  107. [108]Difan Zou and Quanquan Gu. On the convergence of Hamiltonian Monte Carlo with stochastic gradients. In International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 13012–13022, 2021.
  108. [109]Difan Zou, Pan Xu, and Quanquan Gu. Stochastic variance-reduced Hamilton Monte Carlo methods. In International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 6028–6037, 2018.arxiv.org/abs/1802.04791
  109. [110]Difan Zou, Pan Xu, and Quanquan Gu. Sampling from non-log-concave distributions via variance-reduced gradient Langevin dynamics. In International Conference on Artificial Intelligence and Statistics, volume 89 of Proceedings of Machine Learning Research, pages 2936–2945, 2019.
  110. [111]Difan Zou, Pan Xu, and Quanquan Gu. Stochastic gradient Hamiltonian Monte Carlo methods with recursive variance reduction. In Advances in Neural Information Processing Systems, volume 32, 2019.
  111. [112]Difan Zou, Pan Xu, and Quanquan Gu. Faster convergence of stochastic gradient Langevin dynamics for non-log-concave sampling. In Conference on Uncertainty in Artificial Intelligence, volume 161 of Proceedings of Machine Learning Research, pages 1152–1162, 2021.

Paper details

vlnc:W_7p5vyrzg
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 2

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Claude Opus 5.5, GPT-6 Astra

View statement
Looking for related papers…
Contents