Pith. sign in

REVIEW 3 major objections 4 minor 139 references

Exact simulation of diffusions and improved algorithms for log-concave sampling

T0 review · 3 major / 4 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 A new sampler for log-concave distributions simulates the underdamped Langevin diffusion exactly by rejection sampling on path space, achieving query complexity $\tilde O(\kappa^{2/3}d^{1/3}\,\mathrm{polylog}(1/\varepsilon))$ and…

desk verdict Strong new ULD sampling bounds via FORS, but the LSI exponents lean on an unproved imported acceleration theorem. read the letter →

arxiv 2608.05022 v1 pith:BND4437S submitted 2026-08-05 cs.DS cs.NAmath.NAmath.PRmath.STstat.TH

Fan Chen , Sinho Chewi , Alexander Rakhlin , Matthew S. Zhang This is my paper
classification cs.DScs.NAmath.NAmath.PRmath.STstat.TH MSC 68Q2565C0560J6060H35
keywords log-concavesamplingunderdampedLangevindiffusionrejectiononpathspaceGirsanovtheoremRényidivergencequerycomplexityHamiltonianMonteCarlomirror
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

This paper proposes a new way to sample from strongly log-concave distributions: simulate the underdamped Langevin diffusion exactly by rejection sampling on path space, using Girsanov's theorem to build unbiased estimators of the density ratio. From a warm start, the FORS-corrected exponential Euler proposal achieves R\'enyi-accurate samples in $\tilde O(\kappa^{2/3}d^{1/3}\,\mathrm{polylog}(1/\varepsilon))$ gradient queries, and with a Hessian-Lipschitz assumption the dimension dependence improves to $d^{1/5}$. These guarantees improve on the $\tilde O(\kappa d^{1/2})$ complexity of Metropolis-adjusted Langevin and the $d^{1/4}$ dimension dependence of Metropolized Hamiltonian Monte Carlo. The paper argues that correcting the law of an entire continuous-time path, rather than filtering single discrete steps, is the right way to turn accelerated continuous-time mixing into query complexity.

What carries the argument

The central object is FORS (first-order rejection sampling), a rejection step on path space: given a proposal path measure $Q$, the exact path measure $P$ has Girsanov density $dP/dQ = \exp\bigl(-\int_0^T\langle\mu_t,dB_t\rangle - \tfrac12\int_0^T\|\mu_t\|^2\,dt\bigr)$, and FORS accepts the final point with probability $\prod_j \mathrm{Clip}_{[0,1]}((B+W(\xi_j;Z))/(2B))$ using an unbiased estimator $W$ built from $O(1)$ random evaluations of the integrands. For the underdamped Langevin diffusion the paper uses the exponential Euler discretization as the base proposal, and a Picard-iteration refinement involving Hessian-vector products for the higher-order result. The one-step FORS error bound (Lemma 2.4) is composed with the accelerated R\'enyi decay of the exact diffusion (Theorem 2.1) to convert short-time simulation accuracy into $K$-step convergence.

What would settle it

For a one-dimensional Gaussian target with known $\kappa$, simulate the exact underdamped Langevin diffusion and the FORS-corrected exponential Euler proposal for short times and measure $R_{3/2}(\hat\nu\|\pi)$ as $T$, $\kappa$, and $d$ vary; the one-step error must match Proposition 3.1's scaling. Then check the $K$-step composition numerically: if reaching $R_{3/2}\le\delta$ requires $K$ substantially larger than $K\gtrsim (T\gamma)^{-1}\log(qR_q(\nu\|\pi)/\delta)$ with the universal constants of Theorem 2.1, the claimed exponents fail.

Watch

Extended reading notes

Core claim

Under $0 \prec \alpha I \preceq \nabla^2 V \preceq \beta I$ with $\kappa=\beta/\alpha$, the FORS-corrected exponential Euler proposal simulates the underdamped Langevin diffusion so that, from a warm start, the output $\hat\nu$ satisfies $R_{3/2}(\hat\nu\|\pi)\le\delta$ using $\tilde O(\kappa^{2/3}d^{1/3}\,\mathrm{polylog}(1/\delta))$ gradient queries in expectation. With an additional Frobenius-Hessian-Lipschitz bound, the two-stage algorithm achieves $R_{w-1/2}(\hat\nu\|\pi)\le\delta^2$ using $\tilde O(\kappa^{1/2}+\kappa_H^{3/5}d^{1/5}+\kappa^{1/3}(\kappa_H+\kappa_H^{3/11})d^{2/11})$ queries. The paper positions these as improvements over the $\tilde O(\kappa d^{1/2})$ complexity of Metropolis-adjusted Langevin sampling and the $d^{1/4}$ dimension dependence of Metropolized Hamiltonian Monte Carlo, and it applies the same path-space rejection scheme to mirror Langevin diffusion and to non-log-concave Fisher information sampling.

Load-bearing premise

All main complexity exponents rely on Theorem 2.1, an imported black box stating that the underdamped Langevin diffusion decays in R\'enyi divergence at the accelerated rate $Cq\exp(-c\sqrt{\alpha}t)$, and if that theorem's constants or warm-start requirements fail for the orders used here, the quoted $\kappa^{2/3}$, $d^{1/5}$, and $d^{2/11}$ dependencies do not follow.

Editorial extensions

If this is right

  • High-accuracy sampling from strongly log-concave and log-smooth targets costs $\tilde O(\kappa^{2/3}d^{1/3}\,\mathrm{polylog}(1/\varepsilon))$ queries, improving over MALA's $\kappa d^{1/2}$ complexity.
  • Under a Frobenius-Hessian-Lipschitz bound, the cost becomes $\tilde O(\kappa^{1/2}+\kappa_H^{3/5}d^{1/5}+\kappa^{1/3}(\kappa_H+\kappa_H^{3/11})d^{2/11})$, beating the $d^{1/4}$ dimension dependence of Metropolized HMC.
  • For Gaussian targets the bound collapses to $\tilde O(\kappa^{1/2}\,\mathrm{polylog}(d,1/\varepsilon))$, which is near-optimal for the Gaussian class.
  • The mirror Langevin application yields high-accuracy (polylogarithmic in $\varepsilon$) sampling with near-linear dimension dependence under the exact mirror Brownian motion assumption.
  • For non-log-concave targets measured by relative Fisher information, the reduction gives $\tilde O(\beta d^{1/3}K_0/\varepsilon^2)$ queries, improving the dimension dependence from $d^{1/2}$ to $d^{1/3}$.

Reading between the lines

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

  • The paper's main exponents rest on an imported R\'enyi decay theorem that it does not prove; verifying or replacing that theorem with explicit constants is the clearest path to making these bounds unconditional.
  • The progression from $d^{1/3}$ to $d^{1/5}$ suggests a derivative-driven ladder of dimension exponents; testing whether $\nabla^k V$ bounds yield $d^{1/(2k+1)}$-type rates would be a natural extension.
  • A practical test of the framework is to replace the Metropolis filter in HMC implementations with this path-space rejection step and compare acceptance rates and effective sample sizes on Gaussian and logistic-regression targets; the paper itself does not report numerics.
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.

Referee Report

3 major / 4 minor

Summary. The paper develops a new algorithmic framework for high-accuracy sampling from log-concave distributions by using rejection sampling on continuous-time path space. For the underdamped Langevin diffusion, it proposes FORS-corrected exponential Euler proposals, and a Picard-iteration proposal under an additional Frobenius-Hessian-Lipschitz condition. The main complexity claims are: under PI or LSI, a Renyi accuracy delta is achieved with O~(kappa^{2/3} d^{1/3} polylog(1/delta)) gradient queries (Theorem 3.2); and under a third-derivative bound, O~(kappa^{1/2} + kappa_H^{3/5} d^{1/5} + kappa^{1/3}(kappa_H + kappa_H^{3/11}) d^{2/11}) queries (Theorem 4.2). The paper also gives applications to the mirror Langevin diffusion and to Fisher-information bounds for non-log-concave sampling. The finite-time FORS error analysis in Sections 3 and 4 and in the appendices is detailed and internally coherent, with explicit parameter conditions; however, the transition from one-step error bounds to the displayed exponents is not self-contained, because it relies on an externally imported accelerated Renyi decay theorem whose precise hypotheses are not verified in the manuscript.

Significance. If the stated assumptions hold, the paper would substantially improve the prior state of the art: the kappa^{2/3} dependence is the first sublinear-in-kappa high-accuracy guarantee in this model, and the d^{1/5} dimension dependence improves over the d^{1/4} rate of MHMC under the same third-derivative smoothness. The path-space rejection idea is conceptually interesting, and the paper supplies careful one-step error lemmas and explicit tradeoffs between the step size, clipping parameter, and number of iterations. The exposition is also honest about the external dependence: no parameters are fitted, and the main exponents are obtained by solving inequalities rather than by tuning a prediction. Nevertheless, the higher-order and LSI-case results are conditional on precise properties of the imported Theorem 2.1 and on a clean statement of the dominance condition in Theorem 4.1, so the significance is real but not fully secured in the present form.

major comments (3)
  1. [Section 2.1, Theorem 2.1] This theorem is the load-bearing black box for every LSI-case complexity claim in the paper. In the proof of Theorem 3.2(ii) the paper first obtains R_{2q}(bnu || P_{KT}nu) <= delta/3 from FORS and then invokes Theorem 2.1 to conclude R_{2q}(P_{KT}nu || pi) <= delta/2; the same pattern is used with orders r and s in Theorem 4.2 and in Proposition D.6. However, the theorem is only stated, not proved, and no precise statement of the universal constants or of how Cq, the threshold t >= C/gamma, and the required warm-start conditions depend on q is given. Since the displayed exponents kappa^{2/3} d^{1/3} and kappa^{1/2} + kappa_H^{3/5} d^{1/5} + kappa^{1/3}(kappa_H + kappa_H^{3/11}) d^{2/11} are obtained by substituting this bound into a finite-time composition, the central claims are conditional on exactly the version of the theorem assumed here. Please either prove the theorem in an appendix or quote a complete statement from Li and Lu (2026) and verify that it applies at the orders 2q, r, and s and at the warm-start levels used in Lemmas C.3/C.4 and Theorem 4.2.
  2. [Section 4.2, Eq. (5), Theorem 4.1] The displayed dominance condition in Theorem 4.1 contains the term kappa^{3/11} kappa_H^{3/11} ed^{2/11} iota^{1/11}, which does not match the target term A3 = kappa^{1/3}(kappa_H + kappa_H^{3/11}) d^{2/11} used in the proof of Theorem 4.2, nor any of the terms in the subsequent balancing paragraph. The proof of Theorem 4.2 substitutes Eq. (5) to conclude that all terms are bounded by A1 + A2 + A3, so this inconsistency makes the main higher-order guarantee unverifiable as written. Restate Eq. (5) with correct exponents and include explicitly the domination calculation that yields the claimed A1 + A2 + A3 bound.
  3. [Section 4.2, Theorem 4.2 and the paragraph after it] The theorem is stated under LSI but assumes an initial distribution with R_w(nu || pi) = O(d log kappa), and the text only notes that such a warm start can be achieved in the strongly log-concave case. Under LSI alone there is no argument that a simple Gaussian initialization has finite Renyi divergence of order w against pi; log-concave LSI measures can have tails that are heavier than Gaussian, in which case the reverse Renyi divergence from a Gaussian initial law can be infinite. The advertised from-a-cold-start statement is therefore not established in the stated LSI setting. Either restrict Theorem 4.2 to the strongly log-concave setting used in Result 2, where the Gaussian warm start is available, or supply an explicit warm-start construction that works under the stated LSI assumption.
minor comments (4)
  1. [Section 4.2, paragraph after Theorem 4.2] The claimed d^{6/5} PI-setting sampler is asserted with the phrase that details are omitted for brevity; since this is a new advertised consequence, either include the proof or clearly label the statement as a conjecture or future work.
  2. [Section 3, Lemma C.3 and Theorem 3.2] The query-count tail bound is justified by an exponential tail bound for the sum of FORS query counts over the K steps and over restarts, but the restart count is random; please spell out the conditioning argument so that the reader can verify the claimed O(K + log(1/p)) with probability at least 1 - p.
  3. [Section 5, Lemma E.1] The sentence that the expectation can be taken under either the proposal measure Q or P is ambiguous; make explicit that the displayed subexponential bound holds uniformly for both path measures, with constants independent of the starting point.
  4. [Section 1 and Section 4.2, Theorem 4.2] The word exact in the title should be qualified, since the FORS output is the law with tilt E_xi Clip_B(W(xi; Z)), which is only epsilon-close to the target path measure; in addition, the statement of Theorem 4.2 says log^2(1/delta) while the proof concludes with a bound involving iota^4, where iota contains log(1/delta); the eO notation absorbs the discrepancy, but the displayed exponent should be made consistent.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the FORS one-step error is composed with external mixing theorems rather than fitted to the claimed output divergences.

full rationale

The paper's derivation chain is not circular in the sense relevant to this review. The one-step FORS error is quantified in Lemma 2.4 and Propositions 3.1 and 4.1, with explicit moment bounds on the estimator W under stated assumptions. The FORS output distribution is defined by Algorithm 1 and the imported parameter-free guarantee in Theorem 2.3, which is from prior work by the authors but is not the target result of this paper and does not include the claimed ULD complexity exponents. Converting one-step Rényi error into K-step guarantees uses the Rényi weak triangle inequality (Lemma C.4) and the externally stated accelerated decay theorems: Theorem 2.1 is attributed to Li and Lu (2026), and Theorem 2.2 to Cao et al. (2023). These are independent results whose assumptions do not include the paper's conclusions. No parameter is fitted to a subset of data, and no claimed complexity bound is defined in terms of the divergence it is supposed to predict. Self-citations to Chen et al. (2026a,b), Zhang et al. (2026), Chewi et al. (2020), and Chewi and Wibisono (2026) appear as background comparisons or as imported lemmas with external assumptions, not as circular definitions or renamed versions of the main results. The mirror Langevin and Fisher information applications similarly compose FORS errors with independent convergence estimates. Accordingly, no circular step can be exhibited from the paper's equations or citations.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claim rests on a small set of imported theorems and explicit smoothness assumptions. There are no data-fitted constants and no newly invented physical or mathematical entities. The free parameters are algorithmic choices fixed by the analysis, not empirical fits.

free parameters (4)
  • friction gamma = Theta(sqrt(alpha))
    Chosen to activate the accelerated decay of ULD in Theorems 2.1 and 2.2. Not fitted to empirical data, but load-bearing for every main bound.
  • FORS clipping parameter B = Theta(1) in [1,2]
    Balances rejection query cost against clipping bias. Appears in all query complexity bounds and in the error lemmas.
  • simulation horizon T = chosen via inequalities such as T^3 <= gamma / (C beta^2 R (q + log(1/delta)))
    Set in the proofs to satisfy Girsanov and exponential moment conditions. Not fitted to data.
  • Euler step h = h = T/N with N chosen via Eq. (3)
    Controls the discretization error of the unbiased estimator. Chosen algorithmically from problem parameters, not from data.
assumptions (6)
  • domain assumption Theorem 2.1 accelerated Renyi decay for ULD
    Imported from Li and Lu (2026); not proved in this manuscript. Used throughout to compose FORS steps and to obtain cold-start guarantees.
  • domain assumption Theorem 2.2 accelerated L2 decay for ULD
    Imported from Cao et al. (2023); used for the PI-based result in Theorem 3.2(i).
  • domain assumption FORS guarantee from Chen et al. (2026a)
    The paper reuses the first-order rejection sampling algorithm and its query bound as a black box, rather than reproving the base theorem.
  • domain assumption Assumption 1: Hessian Lipschitz in Frobenius norm
    Required for Result 2 and Theorem 4.2; stated explicitly in Section 4.2.
  • domain assumption Assumption 2: self-concordant mirror map and exact mirror Brownian simulation
    Required for the mirror Langevin results in Section 5; inherited from Ahn and Chewi (2021).
  • standard math Girsanov theorem and Novikov condition
    Used to define the density ratio dP/dQ and to justify the path-space rejection scheme; the finite-moment verification is sketched in Remark 2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Exact simulation of diffusions and improved algorithms for log-concave sampling." pith.science (2026). https://pith.science/paper/BND4437S

@misc{pith2026260805022,
  author       = {Pith},
  title        = {Pith review of: Exact simulation of diffusions and improved algorithms for log-concave sampling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BND4437S}},
  note         = {Machine review of arXiv:2608.05022}
}
abstract

We study exact simulation of diffusions via rejection sampling on path space using unbiased estimators of the density ratio obtained from Girsanov's theorem. When applied to the underdamped Langevin diffusion, it yields an algorithm for sampling from a strongly log-concave and log-smooth distribution with condition number $\kappa$, in dimension $d$, to accuracy $\varepsilon$ in R\'enyi divergence, in $\widetilde O(\kappa^{2/3} d^{1/3}\,\mathrm{polylog}(1/\varepsilon))$ queries. Under a third derivative bound, the dimension dependence improves to $d^{1/5}$. This improves substantially over the prior state-of-the-art complexity of $\widetilde O(\kappa d^{1/2}\,\mathrm{polylog}(1/\varepsilon))$ for the Metropolis-adjusted Langevin algorithm, and over the $d^{1/4}$ dimension dependence of Metropolized Hamiltonian Monte Carlo under the same third derivative bound. We also present applications to the mirror Langevin diffusion, and for obtaining Fisher information bounds in the non-log-concave case.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

139 extracted references · 63 canonical work pages

  1. [1]

    and Bartlett, Peter L

    Chatterji, Niladri S. and Bartlett, Peter L. and Long, Philip M. , TITLE =. Bernoulli , FJOURNAL =. 2022 , NUMBER =

    2022

  2. [2]

    Joe Benton and Valentin De Bortoli and Arnaud Doucet and George Deligiannidis , booktitle=. Nearly

  3. [3]

    Complexity of randomized algorithms for underdamped

    Cao, Yu and Lu, Jianfeng and Wang, Lihan , journal=. Complexity of randomized algorithms for underdamped

  4. [4]

    2025 , publisher=

    Lu, Cheng and Zhou, Yuhao and Bao, Fan and Chen, Jianfei and Li, Chongxuan and Zhu, Jun , journal=. 2025 , publisher=

    2025

  5. [5]

    Advances in Neural Information Processing Systems , volume=

    Adapting to unknown low-dimensional structures in score-based diffusion models , author=. Advances in Neural Information Processing Systems , volume=

  6. [6]

    arXiv preprint arXiv:2409.18804 , year=

    Convergence of diffusion models under the manifold hypothesis in high-dimensions , author=. arXiv preprint arXiv:2409.18804 , year=

    arXiv

  7. [7]

    Proceedings of Thirty Eighth Conference on Learning Theory , pages=

    Linear convergence of diffusion models under the manifold hypothesis , author=. Proceedings of Thirty Eighth Conference on Learning Theory , pages=. 2025 , editor=

    2025

  8. [8]

    Mathematics of Operations Research , year=

    Denoising diffusion probabilistic models are optimally adaptive to unknown low dimensionality , author=. Mathematics of Operations Research , year=

Show all 139 references
  1. [9]

    Proceedings of Thirty Eighth Conference on Learning Theory , pages=

    Low-dimensional adaptation of diffusion models: convergence in total variation (extended abstract) , author=. Proceedings of Thirty Eighth Conference on Learning Theory , pages=. 2025 , editor=

    2025

  2. [10]

    Dimension-free convergence of diffusion models for approximate

    Li, Gen and Cai, Changxiao and Wei, Yuting , journal=. Dimension-free convergence of diffusion models for approximate

  3. [11]

    Adaptivity and convergence of probability flow

    Tang, Jiaqi and Yan, Yuling , journal=. Adaptivity and convergence of probability flow

  4. [12]

    and Huan, Stephen and Huang, Jerry and Boffi, Nicholas M

    Zhang, Matthew S. and Huan, Stephen and Huang, Jerry and Boffi, Nicholas M. and Chen, Sitan and Chewi, Sinho , journal=. Sublinear iterations can suffice even for

  5. [13]

    arXiv preprint arXiv:2410.13738 , year=

    Instance-dependent convergence theory for diffusion models , author=. arXiv preprint arXiv:2410.13738 , year=

    arXiv

  6. [14]

    Optimal convergence analysis of

    Jiao, Yuchen and Zhou, Yuchen and Li, Gen , journal=. Optimal convergence analysis of

  7. [15]

    Proceedings of the 41st International Conference on Machine Learning , pages =

    Accelerating convergence of score-based diffusion models, provably , author =. Proceedings of the 41st International Conference on Machine Learning , pages =. 2024 , editor =

    2024

  8. [16]

    Stochastic

    Wu, Yuchen and Chen, Yuxin and Wei, Yuting , journal=. Stochastic

  9. [17]

    arXiv preprint arXiv:2410.23285 , year=

    Provable acceleration for diffusion models under minimal assumptions , author=. arXiv preprint arXiv:2410.23285 , year=

    arXiv

  10. [18]

    Convergence analysis of probability flow

    Huang, Daniel Zhengyu and Huang, Jiaoyang and Lin, Zhengjiang , journal=. Convergence analysis of probability flow. 2025 , publisher=

    2025

  11. [19]

    arXiv preprint arXiv:2506.24042 , year=

    Faster diffusion models via higher-order approximation , author=. arXiv preprint arXiv:2506.24042 , year=

    arXiv

  12. [20]

    Fast convergence for high-order

    Huang, Daniel Zhengyu and Huang, Jiaoyang and Lin, Zhengjiang , journal=. Fast convergence for high-order

  13. [21]

    High-order

    Dang, Thanh and Gurbuzbalaban, Mert and Islam, Mohammad Rafiqul and Yao, Nian and Zhu, Lingjiong , journal=. High-order

  14. [22]

    Fast and efficient parallel sampling using higher order

    Mahajan, Jaideep and Zhang, Kaihong and Liang, Feng and Liu, Jingbo , journal=. Fast and efficient parallel sampling using higher order

  15. [23]

    2014 , PAGES =

    Bakry, Dominique and Gentil, Ivan and Ledoux, Michel , TITLE =. 2014 , PAGES =

    2014

  16. [24]

    2026 , NOTE =

    Chewi, Sinho , TITLE =. 2026 , NOTE =

    2026

  17. [25]

    Concentration inequalities , NOTE =

    Boucheron, St\'. Concentration inequalities , NOTE =. 2013 , PAGES =

    2013

  18. [26]

    and Wibisono, Andre , booktitle =

    Vempala, Santosh S. and Wibisono, Andre , booktitle =. Rapid convergence of the unadjusted. 2019 , publisher =

    2019

  19. [27]

    and Li, Mufan (Bill) and Shen, Ruoqi and Zhang, Matthew S

    Chewi, Sinho and Erdogdu, Murat A. and Li, Mufan (Bill) and Shen, Ruoqi and Zhang, Matthew S. , TITLE =. Found. Comput. Math. , FJOURNAL =. 2025 , NUMBER =

    2025

  20. [28]

    Proceedings of Thirty Sixth Conference on Learning Theory , pages =

    Improved dimension dependence of a proximal algorithm for sampling , author =. Proceedings of Thirty Sixth Conference on Learning Theory , pages =. 2023 , editor =

    2023

  21. [29]

    Conforti, Giovanni and Durmus, Alain and Gentiloni Silveri, Marta , TITLE =. SIAM J. Math. Data Sci. , FJOURNAL =. 2025 , NUMBER =

    2025

  22. [30]

    Gen Li and Yuling Yan , booktitle=

  23. [31]

    Nishant Jain and Tong Zhang , booktitle=. A sharp

  24. [32]

    Jayram, T. S. , TITLE =. Approximation, randomization, and combinatorial optimization , SERIES =

  25. [33]

    and Han, Yanjun and Qian, Jian and Rakhlin, Alexander , booktitle =

    Foster, Dylan J. and Han, Yanjun and Qian, Jian and Rakhlin, Alexander , booktitle =. Online estimation via offline estimation: an information-theoretic framework , volume =

  26. [34]

    Proceedings of Thirty Fifth Conference on Learning Theory , pages =

    Improved analysis for a proximal algorithm for sampling , author =. Proceedings of Thirty Fifth Conference on Learning Theory , pages =. 2022 , editor =

    2022

  27. [35]

    Structured logconcave sampling with a restricted

    Lee, Yin Tat and Shen, Ruoqi and Tian, Kevin , booktitle =. Structured logconcave sampling with a restricted. 2021 , editor =

    2021

  28. [36]

    arXiv preprint arXiv:2601.10708 , year=

    High-accuracy and dimension-free sampling with diffusions , author=. arXiv preprint arXiv:2601.10708 , year=

  29. [37]

    Proceedings of Thirty Ninth Conference on Learning Theory , pages=

    Fast score-based sampling via log-concave reductions , author=. Proceedings of Thirty Ninth Conference on Learning Theory , pages=. 2026 , editor=

    2026

  30. [38]

    Advances in Neural Information Processing Systems , volume=

    Reverse transition kernel: a flexible framework to accelerate diffusion inference , author=. Advances in Neural Information Processing Systems , volume=

  31. [39]

    Keane, M. S. and O'Brien, George L. , journal=. A. 1994 , publisher=

    1994

  32. [40]

    The Annals of Applied Probability , volume=

    Fast simulation of new coins from old , author=. The Annals of Applied Probability , volume=

  33. [41]

    and Chewi, Sinho , title =

    Altschuler, Jason M. and Chewi, Sinho , title =. J. ACM , month =. 2024 , issue_date =

    2024

  34. [42]

    Lu, Jianfeng and Wang, Lihan , TITLE =. Stat. Comput. , FJOURNAL =. 2022 , NUMBER =

    2022

  35. [43]

    The Eleventh International Conference on Learning Representations , year=

    Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions , author=. The Eleventh International Conference on Learning Representations , year=

  36. [44]

    Proceedings of the 34th International Conference on Algorithmic Learning Theory , pages =

    Convergence of score-based generative modeling for general data distributions , author =. Proceedings of the 34th International Conference on Algorithmic Learning Theory , pages =. 2023 , editor =

    2023

  37. [45]

    The probability flow

    Chen, Sitan and Chewi, Sinho and Lee, Holden and Li, Yuanzhi and Lu, Jianfeng and Salim, Adil , booktitle =. The probability flow

  38. [46]

    Lecture notes on statistics and information theory , author =

  39. [47]

    and Zhu, Lingjiong , journal=

    Gao, Xuefeng and Nguyen, Hoang M. and Zhu, Lingjiong , journal=

  40. [48]

    International Conference on Machine Learning , pages=

    Improved analysis of score-based generative modeling: user-friendly bounds under minimal smoothness assumptions , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  41. [49]

    The Twelfth International Conference on Learning Representations , year=

    Towards non-asymptotic convergence for diffusion-based generative models , author=. The Twelfth International Conference on Learning Representations , year=

  42. [50]

    A sharp convergence theory for the probability flow

    Li, Gen and Wei, Yuting and Chi, Yuejie and Chen, Yuxin , journal=. A sharp convergence theory for the probability flow

  43. [51]

    Convergence analysis for general probability flow

    Gao, Xuefeng and Zhu, Lingjiong , booktitle=. Convergence analysis for general probability flow. 2025 , editor=

    2025

  44. [52]

    The Thirteenth International Conference on Learning Representations , year=

    Faster diffusion sampling with randomized midpoints: sequential and parallel , author=. The Thirteenth International Conference on Learning Representations , year=

  45. [53]

    and Ravikumar, Pradeep and Wainwright, Martin J

    Agarwal, Alekh and Bartlett, Peter L. and Ravikumar, Pradeep and Wainwright, Martin J. , TITLE =. IEEE Trans. Inform. Theory , FJOURNAL =. 2012 , NUMBER =

    2012

  46. [54]

    , TITLE =

    Andrieu, Christophe and Roberts, Gareth O. , TITLE =. Ann. Statist. , FJOURNAL =. 2009 , NUMBER =

    2009

  47. [55]

    Proceedings of the 28th International Conference on Machine Learning , pages =

    Welling, Max and Teh, Yee-Whye , title =. Proceedings of the 28th International Conference on Machine Learning , pages =. 2011 , publisher =

    2011

  48. [56]

    Nemeth, Christopher and Fearnhead, Paul , TITLE =. J. Amer. Statist. Assoc. , FJOURNAL =. 2021 , NUMBER =

    2021

  49. [57]

    Wu, Tung-Yu and Wang, Y. X. Rachel and Wong, Wing H. , TITLE =. J. Amer. Statist. Assoc. , FJOURNAL =. 2022 , NUMBER =

    2022

  50. [58]

    Feder and De Sa, Christopher M

    Zhang, Ruqi and Cooper, A. Feder and De Sa, Christopher M. , booktitle =. Asymptotically optimal exact minibatch

  51. [59]

    An efficient minibatch acceptance test for

    Daniel Seita and Xinlei Pan and Haoyu Chen and John Canny , booktitle =. An efficient minibatch acceptance test for. 2018 , month =

    2018

  52. [60]

    Operations research , volume=

    Non-stationary stochastic optimization , author=. Operations research , volume=. 2015 , publisher=

    2015

  53. [61]

    , author=

    Optimal algorithms for online convex optimization with multi-point bandit feedback. , author=. Colt , pages=

  54. [62]

    Advances in Neural Information Processing Systems , volume=

    Bandit convex optimization: towards tight bounds , author=. Advances in Neural Information Processing Systems , volume=

  55. [63]

    IEEE Transactions on Information Theory , volume=

    Information-based complexity, feedback and dynamics in convex programming , author=. IEEE Transactions on Information Theory , volume=. 2011 , publisher=

    2011

  56. [64]

    2019 , journal=

    Simple and optimal high-probability bounds for strongly-convex stochastic gradient descent , author=. 2019 , journal=

    2019

  57. [65]

    arXiv preprint arXiv:2602.01338 , year=

    High-accuracy sampling for diffusion models and log-concave distributions , author=. arXiv preprint arXiv:2602.01338 , year=

    arXiv

  58. [66]

    and Karagulyan, Avetik , TITLE =

    Dalalyan, Arnak S. and Karagulyan, Avetik , TITLE =. Stochastic Process. Appl. , FJOURNAL =. 2019 , NUMBER =

    2019

  59. [67]

    , booktitle =

    Dalalyan, Arnak S. , booktitle =. Further and stronger analogy between sampling and optimization:. 2017 , editor =

    2017

  60. [68]

    Analysis of

    Durmus, Alain and Majewski, Szymon and Miasojedow, B. Analysis of. J. Mach. Learn. Res. , FJOURNAL =. 2019 , PAGES =

    2019

  61. [69]

    Optimal dimension dependence of the

    Chewi, Sinho and Lu, Chen and Ahn, Kwangjun and Cheng, Xiang and Gouic, Thibaut Le and Rigollet, Philippe , booktitle =. Optimal dimension dependence of the. 2021 , editor =

    2021

  62. [70]

    Journal of Machine Learning Research , year =

    Keru Wu and Scott Schmidler and Yuansi Chen , title =. Journal of Machine Learning Research , year =

  63. [71]

    and Salim, Adil and Zhang, Matthew S

    Balasubramanian, Krishna and Chewi, Sinho and Erdogdu, Murat A. and Salim, Adil and Zhang, Matthew S. , booktitle =. Towards a theory of non-log-concave sampling: first-order stationarity guarantees for. 2022 , editor =

    2022

  64. [72]

    Proceedings of Thirty Fifth Conference on Learning Theory , pages =

    The query complexity of sampling from strongly log-concave distributions in one dimension , author =. Proceedings of Thirty Fifth Conference on Learning Theory , pages =. 2022 , editor =

    2022

  65. [73]

    Query lower bounds for log-concave sampling , year=

    Chewi, Sinho and De Dios Pont, Jaume and Li, Jerry and Lu, Chen and Narayanan, Shyam , booktitle=. Query lower bounds for log-concave sampling , year=

  66. [74]

    2025 , journal=

    Mean square error analysis of stochastic gradient and variance-reduced sampling algorithms , author=. 2025 , journal=

    2025

  67. [75]

    Bayesian time series models , PAGES =

    Papaspiliopoulos, Omiros , TITLE =. Bayesian time series models , PAGES =

  68. [76]

    Stochastic Anal

    Wagner, Wolfgang , TITLE =. Stochastic Anal. Appl. , FJOURNAL =. 1988 , NUMBER =

    1988

  69. [77]

    and Bartlett, Peter L

    Mou, Wenlong and Ma, Yi-An and Wainwright, Martin J. and Bartlett, Peter L. and Jordan, Michael I. , journal=. High-order

  70. [78]

    Efficient constrained sampling via the mirror-

    Ahn, Kwangjun and Chewi, Sinho , journal=. Efficient constrained sampling via the mirror-

  71. [79]

    Optimum bounds for the distributions of martingales in

    Pinelis, Iosif , journal=. Optimum bounds for the distributions of martingales in

  72. [80]

    Proceedings of Thirty Ninth Conference on Learning Theory , pages=

    High-accuracy log-concave sampling with stochastic queries , author=. Proceedings of Thirty Ninth Conference on Learning Theory , pages=. 2026 , editor=

    2026

  73. [81]

    and Altschuler, Jason M

    Zhang, Matthew S. and Altschuler, Jason M. and Chewi, Sinho , journal=. Algorithmic warm starts for

  74. [82]

    Exponential ergodicity of mirror-

    Chewi, Sinho and Le Gouic, Thibaut and Lu, Chen and Maunu, Tyler and Rigollet, Philippe and Stromme, Austin , journal=. Exponential ergodicity of mirror-

  75. [83]

    Hanson--

    Adamczak, Rados. Hanson--. Annales de l'Institut Henri Poincar

  76. [84]

    Complexity of non-log-concave sampling in

    Chewi, Sinho and Wibisono, Andre , journal=. Complexity of non-log-concave sampling in

  77. [85]

    The Annals of Applied Probability , volume=

    Exact simulation of diffusions , author=. The Annals of Applied Probability , volume=

  78. [86]

    and Chewi, Sinho and Zhang, Matthew S

    Altschuler, Jason M. and Chewi, Sinho and Zhang, Matthew S. , booktitle=. Shifted composition

  79. [87]

    Handbook of

    Brooks, Steve and Gelman, Andrew and Jones, Galin and Meng, Xiao-Li , year=. Handbook of

  80. [88]

    Wasserstein control of mirror

    Zhang, Kelvin Shuangjian and Peyr. Wasserstein control of mirror. Conference on learning theory , pages=. 2020 , organization=

    2020

  81. [89]

    Jiang, Qijia , journal=. Mirror

  82. [90]

    and Wibisono, Andre , booktitle=

    Li, Ruilin and Tao, Molei and Vempala, Santosh S. and Wibisono, Andre , booktitle=. The mirror. 2022 , organization=

    2022

  83. [91]

    , journal=

    Gatmiry, Khashayar and Vempala, Santosh S. , journal=. Convergence of the

  84. [92]

    and Yu, Bin , journal=

    Dwivedi, Raaz and Chen, Yuansi and Wainwright, Martin J. and Yu, Bin , journal=. Log-concave sampling:

  85. [93]

    and Rosenthal, Jeffrey S

    Roberts, Gareth O. and Rosenthal, Jeffrey S. , journal=. Optimal scaling of discrete approximations to. 1998 , publisher=

    1998

  86. [94]

    and Yu, Bin , journal=

    Chen, Yuansi and Dwivedi, Raaz and Wainwright, Martin J. and Yu, Bin , journal=. Fast mixing of

  87. [95]

    International Conference on Algorithmic Learning Theory , pages=

    Fisher information lower bounds for sampling , author=. International Conference on Algorithmic Learning Theory , pages=. 2023 , organization=

    2023

  88. [96]

    and Roberts, Gareth O

    Beskos, Alexandros and Pillai, Natesh S. and Roberts, Gareth O. and Sanz-Serna, Jes. Optimal tuning of the hybrid. Bernoulli , volume=

  89. [97]

    , journal=

    Neal, Radford M. , journal=. 2011 , publisher=

    2011

  90. [98]

    A simple proof of the mixing of

    Chen, Yuansi and Gatmiry, Khashayar , journal=. A simple proof of the mixing of

  91. [99]

    When does

    Yuansi Chen and Khashayar Gatmiry and Minhui Jiang , year=. When does

  92. [100]

    Advances in Neural Information Processing Systems , volume=

    The randomized midpoint method for log-concave sampling , author=. Advances in Neural Information Processing Systems , volume=

  93. [101]

    The Fourteenth International Conference on Learning Representations , year=

    Poisson midpoint method for log-concave sampling: beyond the strong error lower bounds , author=. The Fourteenth International Conference on Learning Representations , year=

  94. [102]

    Randomized

    Bou-Rabee, Nawaf and Kleppe, Tore , journal=. Randomized

  95. [103]

    Bernoulli , volume=

    Retrospective exact simulation of diffusion sample paths with applications , author=. Bernoulli , volume=. 2006 , publisher=

    2006

  96. [104]

    Kakkad, Dwija and Vats, Dootika , journal=. Exact

  97. [105]

    Gon. Exact. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2023 , publisher=

    2023

  98. [106]

    Brazilian Journal of Probability and Statistics , volume=

    Gon. Brazilian Journal of Probability and Statistics , volume=

  99. [107]

    Efficient

    Vats, Dootika and Gon. Efficient. Biometrika , volume=. 2022 , publisher=

    2022

  100. [108]

    Sampling as optimization in the space of measures: the

    Wibisono, Andre , booktitle =. Sampling as optimization in the space of measures: the. 2018 , editor =

    2018

  101. [109]

    Jordan, Richard and Kinderlehrer, David and Otto, Felix , TITLE =. SIAM J. Math. Anal. , FJOURNAL =. 1998 , NUMBER =

    1998

  102. [110]

    , TITLE =

    Nesterov, Yurii E. , TITLE =. Dokl. Akad. Nauk SSSR , FJOURNAL =. 1983 , NUMBER =

    1983

  103. [111]

    and Parrilo, Pablo A

    Altschuler, Jason M. and Parrilo, Pablo A. , TITLE =. Math. Program. , FJOURNAL =. 2025 , NUMBER =

    2025

  104. [112]

    , TITLE =

    Su, Weijie and Boyd, Stephen and Cand\`es, Emmanuel J. , TITLE =. J. Mach. Learn. Res. , FJOURNAL =. 2016 , PAGES =

    2016

  105. [113]

    Cao, Yu and Lu, Jianfeng and Wang, Lihan , TITLE =. Arch. Ration. Mech. Anal. , FJOURNAL =. 2023 , NUMBER =

    2023

  106. [114]

    Probability Theory and Related Fields , volume=

    Non-reversible lifts of reversible diffusion processes and relaxation times , author=. Probability Theory and Related Fields , volume=. 2026 , publisher=

    2026

  107. [115]

    Chen, Fang and Lov\'asz, L\'aszl\'o. Lifting. Annual

  108. [116]

    , TITLE =

    Diaconis, Persi and Holmes, Susan and Neal, Radford M. , TITLE =. Ann. Appl. Probab. , FJOURNAL =. 2000 , NUMBER =

    2000

  109. [117]

    A sharp hypocoercive entropy decay estimate for underdamped

    Jianfeng Lu , year=. A sharp hypocoercive entropy decay estimate for underdamped

  110. [118]

    and Chewi, Sinho , journal=

    Altschuler, Jason M. and Chewi, Sinho , journal=. Shifted composition. 2025 , volume=

    2025

  111. [119]

    Altschuler and Sinho Chewi , title=

    Jason M. Altschuler and Sinho Chewi , title=. Found. Comput. Math. , FJOURNAL =

  112. [120]

    Complexity of log-concave sampling and applications

    Kook, Yunbum. Complexity of log-concave sampling and applications. 2026

    2026

  113. [121]

    The Journal of Chemical Physics , volume=

    Equation of state calculations by fast computing machines , author=. The Journal of Chemical Physics , volume=. 1953 , publisher=

    1953

  114. [122]

    Keith , TITLE =

    Hastings, W. Keith , TITLE =. Biometrika , FJOURNAL =. 1970 , NUMBER =

    1970

  115. [123]

    Space-time log-

    Bowen Li and Jianfeng Lu , year=. Space-time log-

  116. [124]

    2026 , journal=

    Windowed thinning and query complexity for the bouncy particle and zigzag samplers , author=. 2026 , journal=

    2026

  117. [125]

    2026 , journal=

    On the entropic convergence for piecewise deterministic samplers: speedup and obstruction , author=. 2026 , journal=

    2026

  118. [126]

    Hypocoercivity , JOURNAL =

    Villani, C\'. Hypocoercivity , JOURNAL =. 2009 , NUMBER =

    2009

  119. [127]

    Albritton, Dallas and Armstrong, Scott and Mourrat, Jean-Christophe and Novack, Matthew , TITLE =. Anal. PDE , FJOURNAL =. 2024 , NUMBER =

    2024

  120. [128]

    Brigati, Giovanni , TITLE =. Kinet. Relat. Models , FJOURNAL =. 2023 , NUMBER =

    2023

  121. [129]

    Convergence of non-reversible

    Andreas Eberle and Arnaud Guillin and Leo Hahn and Francis L\". Convergence of non-reversible. 2025 , journal=

    2025

  122. [130]

    Space-time divergence lemmas and optimal non-reversible lifts of diffusions on

    Eberle, Andreas and L. Space-time divergence lemmas and optimal non-reversible lifts of diffusions on. Journal of Functional Analysis , volume=

  123. [131]

    Lehec, Joseph , TITLE =. Math. Stat. Learn. , FJOURNAL =. 2025 , NUMBER =

    2025

  124. [132]

    2025 , journal=

    Quantitative hypocoercivity and lifting of classical and quantum dynamics , author=. 2025 , journal=

    2025

  125. [133]

    Lu, Jianfeng and Wang, Lihan , TITLE =. Ann. Appl. Probab. , FJOURNAL =. 2022 , NUMBER =

    2022

  126. [134]

    Sharp hypocoercive convergence estimates for underdamped

    Zexi Fan and Bowen Li and Jianfeng Lu , year=. Sharp hypocoercive convergence estimates for underdamped

  127. [135]

    Nesterov acceleration for the

    Pierre Monmarché , year=. Nesterov acceleration for the

  128. [136]

    Accelerated mixing time of randomized

    Siddharth Mitra and Vishwak Srinivasan and Xiuyuan Wang and Andre Wibisono , year=. Accelerated mixing time of randomized

  129. [137]

    Logsmooth gradient concentration and tighter runtimes for

    Lee, Yin Tat and Shen, Ruoqi and Tian, Kevin , booktitle =. Logsmooth gradient concentration and tighter runtimes for. 2020 , editor =

    2020

  130. [138]

    2026 , journal=

    A proximal gradient algorithm for composite log-concave sampling , author=. 2026 , journal=

    2026

  131. [139]

    Fast sampling from constrained spaces using the

    Srinivasan, Vishwak and Wibisono, Andre and Wilson, Ashia , booktitle =. Fast sampling from constrained spaces using the. 2024 , editor =

    2024

Pith tools

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