Pith. sign in

REVIEW 4 major objections 5 minor 29 references

Metropolis-adjusted Subdifferential Langevin Algorithm

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read MASLA, a subgradient-based Metropolis-adjusted Langevin sampler, is stationary and reversible for nonsmooth nonconvex targets under mild regularity.

desk verdict A correct but modest extension of MALA to nonsmooth targets; the stationarity proof has a fixable flaw, and the paper overstates convergence support. read the letter →

arxiv 2507.06950 v1 pith:ORE23O6I submitted 2025-07-09 stat.ME

Ning Ning This is my paper · ORCID
classification stat.ME MSC 60J2262F1565C0549J52
keywords Metropolis-adjustedLangevinalgorithmconservativefieldssubgradientMCMCnonsmoothpotentialsnonconvexsamplingpathdifferentiabilityreversibilityMarkovchainMonteCarlo
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 introduces MASLA, a Markov chain Monte Carlo method that replaces the gradient in the Metropolis-adjusted Langevin algorithm with an arbitrary element of a conservative field—a set-valued generalized derivative that agrees with the gradient almost everywhere and obeys a chain rule. The central claim is that this substitution preserves the Metropolis-Hastings guarantee: under a mild regularity condition (the potential is twice continuously differentiable at almost every point, and the chain starts from an absolutely continuous distribution), the chain has the target as its stationary distribution and is reversible for almost every step size. This matters because MALA's differentiability requirement rules out many practical targets, such as Laplace priors, total-variation regularized models, and ReLU-network potentials, which are locally Lipschitz but nonsmooth and often nonconvex. The paper demonstrates that MASLA matches the convergence of state-of-the-art proximal and subgradient Langevin methods where those apply, and that it can sample a nonconvex potential whose proximal operator is not single-valued, where proximal-based samplers are not directly usable.

What carries the argument

The machinery is the conservative (set-valued) field $D_U$, a generalized derivative that is closed-graph, nonempty compact-valued, and characterized by a vanishing loop integral; it coincides with $\{\nabla U\}$ almost everywhere and admits a chain rule for absolutely continuous curves. The Clarke subgradient is the minimal convex conservative field, so any function with a conservative field—convex, concave, Clarke regular, semialgebraic, or tame functions—falls in scope. MASLA's proposal draws $\beta(X_n) \in D_U(X_n)$ and uses it exactly where MALA would use $\nabla U(X_n)$; the chain rule and the almost-everywhere agreement with the gradient are what let the density-existence argument survive at points where $U$ is not differentiable.

What would settle it

Run MASLA on a locally Lipschitz, path-differentiable potential whose nondifferentiability set has positive Lebesgue measure, with an absolutely continuous start and several step sizes, and test whether the empirical stationary distribution equals the target and detailed balance holds; a positive-measure set of step sizes for which reversibility fails would show the $C^{2}$-almost-everywhere assumption is load-bearing rather than technical.

Watch

Extended reading notes

Core claim

The paper's central claim is that the Metropolis-Hastings correction rescues a subgradient-based Euler proposal from discretization bias even when the target's log-density is nonsmooth and nonconvex. Theorem 2.13 states that if the potential $U$ is twice continuously differentiable at $\lambda$-almost every point and the initial distribution is absolutely continuous with respect to Lebesgue measure, then for almost every step size $\gamma$ the MASLA chain has $\pi$ as its stationary distribution and is reversible with respect to $\pi$. The proof shows the proposal map $g_{\gamma,z}(x) = (x - \gamma\nabla U(x))\mathbf{1}_{D_U}(x) + (2\gamma)^{1/2}z$ is a local diffeomorphism except on a set of step sizes that is Lebesgue-null, which guarantees the proposal has a density and detailed balance can be verified. This makes MASLA a valid MCMC sampler for locally Lipschitz, generally non-differentiable, non-convex targets, provided the potential admits a conservative field.

Load-bearing premise

The guarantee rests on the target's log-density being twice continuously differentiable at almost every point and admitting a conservative field; if a nonsmooth target fails either condition, the proposal may have no density or the local-diffeomorphism argument collapses.

Editorial extensions

If this is right

  • MASLA provides a valid MCMC algorithm for targets whose log-densities are locally Lipschitz, generally non-differentiable, and non-convex, as long as they are path-differentiable and C^2 almost everywhere.
  • The Metropolis correction removes the discretization bias: in the paper's comparison on f(x)=|x^2-1|, MASLA reaches a TV error of 0.014 versus 0.117 for the unadjusted subgradient chain, and a Wasserstein error of 0.008 versus 0.092.
  • On the TV-L2 target, MASLA's Wasserstein and TV convergence rates are comparable to proximal-subgradient Langevin algorithms across step sizes 10^{-3}, 10^{-4}, 10^{-5}, so the generalization to nonsmooth potentials does not sacrifice the accuracy of MALA-type samplers.
  • The proposal only needs an element of a conservative field, which automatic differentiation can supply for compositions of smooth maps and activations; this connects MASLA to neural-network and other nonsmooth learning targets.

Reading between the lines

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

  • The theorem's 'almost every step size' guarantee leaves open which specific step sizes are safe; the proof suggests that step sizes making a certain determinant vanish on a positive-measure set are the exceptions.
  • The C^2-almost-everywhere assumption is stronger than path differentiability alone; extending the density-existence proof to potentials whose nondifferentiability set has positive measure would be a genuine extension beyond this paper.
  • Because MASLA selects an arbitrary element of the conservative field, the chain's trajectory may depend on that choice; whether a specific selection rule, such as the minimal-norm subgradient, systematically improves mixing is an open question the paper does not test.
  • The single numerical example with a nonconvex, non-prox-able potential suggests MASLA could fill a gap for Bayesian models with nonconvex nonsmooth priors, but only one such case is demonstrated, so broader empirical evidence is the next step.
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

4 major / 5 minor

Summary. The paper introduces the Metropolis-Adjusted Subdifferential Langevin Algorithm (MASLA), a Metropolis-Hastings correction of a Langevin-type proposal in which the gradient is replaced by an element of a conservative (Clarke-type) subdifferential field of a locally Lipschitz, possibly nonconvex potential U. Section 2 states Theorem 2.13, which claims that, under a C^2-almost-everywhere condition and an absolutely continuous initial law, the MASLA chain has π as a stationary distribution and is reversible with respect to π for almost every step size. Section 3 compares MASLA numerically with Grad-sub, Prox-sub, P-MALA, and MYULA on a TV-L2 target, and with USLA on the nonsmooth example |x^2-1|.

Significance. If the stationarity claim is valid, MASLA is a natural and potentially useful extension of MALA to nonsmooth, nonconvex targets, and the connection to conservative fields (Bolte and Pauwels, 2021) is appropriate. The numerical experiments follow an existing benchmark and are suggestive of competitive performance. However, the proof of the central theorem as written is invalid, and the paper does not establish the 'Exponential convergence' listed in the keywords; convergence of the sampler is only demonstrated empirically. The manuscript therefore needs a corrected proof and a clarified or weakened set of claims before the central contribution can be accepted.

major comments (4)
  1. [Section 2.2, proof of Theorem 2.13] The displayed inequality ρeQγ(A) = ρ⊗Φ({(x,z): g_{γ,z}(x)∈A}) ≤ λ⊗Φ({(x,z): g_{γ,z}(x)∈A}) does not follow from ρ ≪ λ. Absolute continuity gives ρ(B)=0 whenever λ(B)=0; it does not give ρ(B) ≤ λ(B) for arbitrary B (e.g., the uniform density on [0,1/2] is 2). Consequently the argument does not prove that ρeQγ ≪ λ, and hence does not prove the existence of the proposal density qγ on which the detailed-balance calculation in (2.8)–(2.9) depends. This is a load-bearing gap in the stationarity theorem. The theorem can be repaired more simply: for any measurable selection β, the proposal kernel has density qγ(x,y) = (4πγ)^{-d/2} exp(−||y−x+γβ(x)||^2/(4γ)), so if the law of Xn has a Lebesgue density then the law of the proposal does as well. The local-diffeomorphism/C²-a.e. apparatus is unnecessary; I recommend replacing this part of the proof with the Gaussian-density argument.
  2. [Theorem 2.13 statement; Section 1.2, eq. (1.8)] Theorem 2.13 omits the path-differentiability assumption that is actually used by the algorithm. The proposal (1.8)–(2.6) requires a conservative field DU and a selection β(x) ∈ DU(x); without path differentiability β may be undefined, so the theorem's hypotheses do not guarantee that the algorithm is well defined. The assumed C²-a.e. condition is also much stronger than needed, and the theorem does not state that U admits a conservative field. The theorem should explicitly assume that U is path differentiable (or that DU is a conservative field for U) and that β is a measurable selection. In that case the Gaussian-density argument shows the stationarity/reversibility conclusion holds for every γ > 0, not merely almost every γ.
  3. [Abstract, keywords; Section 3.1.2] The manuscript claims 'Exponential convergence' in its keywords and states in Section 3.1.2 that 'All algorithms exhibit exponential ergodicity to the target density.' No theorem in Section 2 establishes convergence of the MASLA chain to π from an arbitrary initial distribution, let alone an exponential rate; Theorem 2.13 only asserts stationarity and reversibility. The Wasserstein and TV plots in Figures 4–5 are finite-horizon empirical distances, which do not constitute a convergence proof. Please either add a convergence theorem (e.g., geometric ergodicity under appropriate conditions) or remove or qualify the exponential-convergence and exponential-ergodicity claims.
  4. [Section 1.2, eq. (1.8); proof of Theorem 2.13] The algorithm description says MASLA 'draws a subgradient from the set-valued field DU(x) at each iteration,' but the stationarity proof treats β as a fixed deterministic function via (2.7). If β is random when DU(x) has more than one element, the proposal transition is a Gaussian mixture and the acceptance probability in (2.9) must use the mixture density, not the density of the Gaussian with the realized β. The paper should specify whether β is a fixed measurable selection or a random draw, and adjust the proof (or the implementation description) accordingly. In practice the ambiguity only matters on a Lebesgue-null set, given Proposition 2.8, but the proof should state this explicitly.
minor comments (5)
  1. [Throughout] There are several typographical errors, including 'discartion' in Section 3.1.1, 'Speciifically' in Section 2.2, 'satifies' in the proof of Theorem 2.13, and 'Proposition 2.81' where Proposition 2.8 is meant.
  2. [Section 3.1.3] In the sentence 'Experiments were conducted with step sizes τ ∈ 10^{-3}, 10^{-4}, 10^{-5},' the set notation is missing braces and should be τ ∈ {10^{-3}, 10^{-4}, 10^{-5}}.
  3. [Figure 4 caption] The caption says 'dashed lines represent theoretical convergence rates' for the competitors, but no theoretical rate is plotted or derived for MASLA; please clarify what the dashed lines represent for each method.
  4. [Section 3.1.1] The phrase 'All experiments in this section has no burn-in discartion' is grammatically unclear; it should be rewritten, for example as 'No burn-in is discarded in the experiments in this section.'
  5. [Section 2.2, proof of Theorem 2.13] The definition of eQγ in (2.10)–(2.11) uses a simplified proposal with drift ∇U(x)1_{D_U}(x); the proof should explicitly note that this simplification is equal to the actual proposal almost surely under the absolutely continuous law, which is what (2.14) is intended to convey.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: stationarity is obtained by the standard MH detailed-balance argument, and the external citations are not load-bearing self-citation.

full rationale

The derivation of stationarity in Theorem 2.13 is the standard Metropolis–Hastings detailed-balance argument: once the proposal density qγ exists, the accept/reject step makes π reversible by construction. The existence of qγ is argued from the Gaussian form of the proposal; the proof's local-diffeomorphism/absolute-continuity step contains a mathematical gap (the inequality ρ eQγ(A) ≤ λ⊗Φ(...) does not follow from ρ ≪ λ), but that is a correctness flaw, not circularity. The imported results from Bolte and Pauwels (2021) (conservative fields, chain rule, D = {∇f} a.e.) are cited external facts with stated assumptions, and the sole self-citation (Ning 2025) is tangential support for a known optimal-scaling constant in the introduction, not load-bearing for MASLA's stationarity or numerical results. The numerical experiments benchmark against external code and targets (Habring et al. 2024), so there is no fitted parameter renamed as a prediction. No step of the derivation reduces to its own input by definition.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The central result rests on imported results on conservative fields (Bolte and Pauwels 2021) and an implicit path-differentiability assumption; no free parameters are fitted to data. The step size γ is a standard tuning parameter and the stationarity result holds for almost every γ.

assumptions (3)
  • standard math Conservative fields satisfy Proposition 2.8: D = {∇f} a.e., ∂c f is a conservative field, and criticality conditions hold.
    Imported from Bolte and Pauwels (2021) without proof; used in Theorem 2.13 to assert P(Xn ∉ DU) = 0.
  • domain assumption The target U admits a conservative field DU (path differentiability).
    Needed so the proposal β(Xn) ∈ DU(Xn) is well-defined; not listed as an assumption in Theorem 2.13 but used throughout.
  • standard math Clarke subgradient is a conservative field for functions satisfying the chain rule (Proposition 2.9).
    Connects the Clarke subgradient to the conservative-field framework; taken from Bolte and Pauwels (2021).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Metropolis-adjusted Subdifferential Langevin Algorithm." pith.science (2026). https://pith.science/paper/ORE23O6I

@misc{pith2026250706950,
  author       = {Pith},
  title        = {Pith review of: Metropolis-adjusted Subdifferential Langevin Algorithm},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ORE23O6I}},
  note         = {Machine review of arXiv:2507.06950}
}
read the original abstract

The Metropolis-Adjusted Langevin Algorithm (MALA) is a widely used Markov Chain Monte Carlo (MCMC) method for sampling from high-dimensional distributions. However, MALA relies on differentiability assumptions that restrict its applicability. In this paper, we introduce the Metropolis-Adjusted Subdifferential Langevin Algorithm (MASLA), a generalization of MALA that extends its applicability to distributions whose log-densities are locally Lipschitz, generally non-differentiable, and non-convex. We evaluate the performance of MASLA by comparing it with other sampling algorithms in settings where they are applicable. Our results demonstrate the effectiveness of MASLA in handling a broader class of distributions while maintaining computational efficiency.

Figures

Figures reproduced from arXiv: 2507.06950 by the authors.

Figure 1
Figure 1. Trajectories of MALA and ULA for sampling from [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Histograms of MALA and ULA samples for π(x) ∝ exp(−|x| 4/4) with step size γ = 0.001 over 100,000 iterations. The left plot shows the MALA his￾togram, basically matching the target density (red curve), indicating ergodicity due to the Metropolis-Hastings correction. The right plot shows the ULA his￾togram, whose samples are transient under small step size. follows: f(y) =    −x − 1, x < −1 x + 1, −1 ≤ x < 0… view at source ↗
Figure 3
Figure 3. Comparison of MASLA-sampled initial and final distributions with [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Comparison of MASLA, Grad-sub (Habring et al., 2024b), and Prox￾sub (Habring et al., 2024b) for TV −L 2 sampling in Wasserstein distance with σ = 1, λ = 5, and step sizes τ ∈ {10−3 , 10−4 , 10−5}. Colors denote different τ values, dashed lines represent theoretical con…
Figure 5
Figure 5. Figure 5: Comparison of MASLA, Prox-sub (Habring et al., 2024b), P-MALA (Pereyra, 2016), and MYULA (Durmus et al., 2022) for TV −L 2 sampling in TV distance with σ = 1, λ = 5, and step sizes τ ∈ {10−3 , 10−4 , 10−5}. Colors represent different τ values. Thick transparent lines d…
Figure 6
Figure 6. Figure 6: Comparison the performance of USLA and MASLA chains with true [PITH_FULL_IMAGE:figures/full_fig_p023_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 28 canonical work pages

  1. [1]

    Bernton, E. (2018). L angevin M onte C arlo and JKO splitting. In Proceedings of the 31st Conference on Learning Theory , volume 75 of Proceedings of Machine Learning Research , pages 1777--1798. PMLR

    work page 2018

  2. [2]

    and Pauwels, E

    Bolte, J. and Pauwels, E. (2021). Conservative set valued fields, automatic differentiation, stochastic gradient methods and deep learning. Mathematical Programming , 188:19--51

    work page 2021

  3. [3]

    E., and Nocedal, J

    Bottou, L., Curtis, F. E., and Nocedal, J. (2018). Optimization methods for large-scale machine learning. SIAM Review , 60(2):223--311

    work page 2018

  4. [4]

    M., Chang, J., Lucas, J

    Carvalho, C. M., Chang, J., Lucas, J. E., Nevins, J. R., Wang, Q., and West, M. (2008). High-dimensional sparse factor modeling: applications in gene expression genomics. Journal of the American Statistical Association , 103(484):1438--1456

    work page 2008

  5. [5]

    Chaari, L., Tourneret, J.-Y., Chaux, C., and Batatia, H. (2016). A H amiltonian M onte C arlo method for non-smooth energy sampling. IEEE Transactions on Signal Processing , 64(21):5585--5594

    work page 2016

  6. [6]

    Clarke, F. H. (1990). Optimization and nonsmooth analysis . SIAM

    work page 1990

  7. [7]

    H., Ledyaev, Y

    Clarke, F. H., Ledyaev, Y. S., Stern, R. J., and Wolenski, P. R. (2008). Nonsmooth analysis and control theory , volume 178. Springer Science & Business Media

    work page 2008

  8. [8]

    Coste, M. (1999). An Introduction to o-minimal geometry . RAAG notes. Institut de Recherche Mathématique de Rennes

    work page 1999

Show all 29 references
  1. [9]

    Dalalyan, A., Grappin, E., and Paris, Q. (2018). On the exponentially weighted aggregate with the L aplace prior. The Annals of Statistics , 46(5):2452--2478

    2018

  2. [10]

    Deng, W., Feng, Q., Gao, L., Liang, F., and Lin, G. (2020). Non-convex learning via replica exchange stochastic gradient MCMC . In International Conference on Machine Learning , pages 2474--2483. PMLR

    2020

  3. [11]

    V., Kungurtsev, V., and Marecek, J

    Difonzo, F. V., Kungurtsev, V., and Marecek, J. (2022). Stochastic Langevin differential inclusions with applications to machine learning. arXiv preprint arXiv:2206.11533

    2022 arXiv

  4. [12]

    Durmus, A., Majewski, S., and Miasojedow, B. (2019). Analysis of Langevin Monte Carlo via convex optimization. Journal of Machine Learning Research , 20(73):1--46

    2019

  5. [13]

    Durmus, A., Moulines, \'E ., and Pereyra, M. (2018). Efficient Bayesian computation by proximal Markov chain Monte Carlo : when Langevin meets Moreau . SIAM Journal on Imaging Sciences , 11(1):473--506

    2018

  6. [14]

    Durmus, A., Moulines, \'E ., and Pereyra, M. (2022). A proximal Markov chain Monte Carlo method for B ayesian inference in imaging inverse problems: When L angevin meets M oreau. SIAM Review , 64(4):991--1028

    2022

  7. [15]

    V., and Shamai, S

    Dytso, A., Bustin, R., Poor, H. V., and Shamai, S. (2018). Analytical properties of generalized G aussian distributions. Journal of Statistical Distributions and Applications , 5:1--40

    2018

  8. [16]

    Evans, L. C. and Gariepy, R. F. (2015). Measure Theory and Fine Properties of Functions . Chapman and Hall/CRC, London, revised edition

    2015

  9. [17]

    Z., Boisbunon, A., Chambon, S., and Chapel, L

    Flamary, R., Courty, N., Gramfort, A., Alaya, M. Z., Boisbunon, A., Chambon, S., and Chapel, L. (2021). POT: Python Optimal Transport . Journal of Machine Learning Research , 22:1--8

    2021

  10. [18]

    Habring, A., Holler, M., and Pock, T. (2024a). Source Code to Reproduce the Results of ``Subgradient Langevin Methods for Sampling from Nonsmooth Potentials'' . https://github.com/habring/subgradient_langevin

    2024

  11. [19]

    Habring, A., Holler, M., and Pock, T. (2024b). Subgradient Langevin methods for sampling from nonsmooth potentials. SIAM Journal on Mathematics of Data Science , 6(4):897--925

    2024

  12. [20]

    Jain, P., Kar, P., et al. (2017). Non-convex optimization for machine learning. Foundations and Trends in Machine Learning , 10(3-4):142--363

    2017

  13. [21]

    Ning, N. (2025). Convergence of dirichlet forms for mcmc optimal scaling with dependent target distributions on large graphs. The Annals of Applied Probability , 35(2):898--935

    2025

  14. [22]

    O'Hara, R. B. and Sillanp \"a \"a , M. J. (2009). A review of B ayesian variable selection methods: what, how and which. Bayesian Analysis , 4(1):85--117

    2009

  15. [23]

    and Casella, G

    Park, T. and Casella, G. (2008). The B ayesian lasso. Journal of the American Statistical Association , 103(482):681--686

    2008

  16. [24]

    Pereyra, M. (2016). Proximal M arkov chain M onte C arlo algorithms. Statistics and Computing , 26(4):745--760

    2016

  17. [25]

    O., Gelman, A., and Gilks, W

    Roberts, G. O., Gelman, A., and Gilks, W. R. (1997). Weak convergence and optimal scaling of random walk M etropolis algorithms. The Annals of Applied Probability , 7(1):110--120

    1997

  18. [26]

    Roberts, G. O. and Rosenthal, J. S. (1998). Optimal scaling of discrete approximations to L angevin diffusions. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , 60(1):255--268

    1998

  19. [27]

    Roberts, G. O. and Tweedie, R. L. (1996). Exponential convergence of Langevin distributions and their discrete approximations . Bernoulli , 2(4):341 -- 363

    1996

  20. [28]

    Seeger, M. W. (2008). Bayesian inference and optimal design for the sparse linear model. Journal of Machine Learning Research , 9:759--813

    2008

  21. [29]

    Van den Dries, L. (1998). Tame topology and o-minimal structures , volume 248. Cambridge university press

    1998

Pith tools

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