{"id":"4aced38c-0e41-444e-8ae4-dd677a321454","arxiv_id":"2507.06950","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"MASLA replaces the gradient in MALA with an element of a conservative subdifferential field, yielding a Metropolis-Hastings sampler that is reversible for locally Lipschitz, non-convex targets when the potential is twice differentiable almost everywhere.","lead":"This paper introduces MASLA, a version of the Metropolis-adjusted Langevin algorithm that uses subgradients instead of gradients so it can sample from targets with non-smooth, non-convex log-densities. The authors prove the sampler is reversible with respect to the target for almost all step sizes and show numerical examples where it matches or beats existing non-smooth samplers.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proof of Theorem 2.13 is invalid: the inequality ρeQγ(A) ≤ λ⊗Φ(...) does not follow from absolute continuity; the stationarity theorem is still true but needs a simpler argument.","rationale":"The paper proposes MASLA, a Metropolis-Hastings sampler using a drift drawn from a conservative field. The central theoretical claim is Theorem 2.13, which asserts stationarity/reversibility for almost every step size under a C^2-a.e. assumption. The reader's verdict is CONDITIONAL, mainly due to the absence of convergence analysis. My stress-test focuses on the proof of Theorem 2.13 itself. The final paragraph of the proof contains a false measure-theoretic inequality: from ρ ≪ λ the authors conclude ρ⊗Φ ≤ λ⊗Φ, which does not follow; densities can exceed 1. This invalidates the proof as written, because the conclusion that ρeQγ(A)=0 for Lebesgue-null A relies on this inequality. The theorem itself is true, but for a much simpler reason: the proposal is a Gaussian convolution, so its conditional density exists for any measurable drift and the MH detailed-balance calculation goes through without any C^2 assumption. Thus the paper's restrictive assumptions and complicated local-diffeomorphism argument are unnecessary, and the proof should be replaced. This is load-bearing because the paper's main theoretical guarantee is not rigorously established as written. The reader's weakest assumption about C^2 a.e. is beside the point; the flaw is the inequality. The overall verdict remains CONDITIONAL: the theorem is likely correct and fixable, but the manuscript needs a corrected proof and should either weaken Theorem 2.13's assumptions or prove convergence before the claims are relied upon.","tokens_in":15468,"tokens_out":27672,"duration_ms":313670,"concrete_test":"Check the false step: take d=1, ρ = Uniform[0,1/2] (density 2), and A = [0,1/4]. Then ρ(A)=1/2 > λ(A)=1/4, so the inequality ρ⊗Φ(B) ≤ λ⊗Φ(B) fails for B=A×R. This confirms the proof's inequality is invalid. Then verify the correct argument: for any measurable β, show the proposal density exists by convolution, and re-derive stationarity via detailed balance; if this succeeds, Theorem 2.13 holds without C^2 assumptions but the paper's proof needs revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.2, final paragraph of the proof of Theorem 2.13: the authors claim that for ρ ≪ λ, ρeQγ(A) = ρ⊗Φ({(x,z): g_{γ,z}(x)∈A}) ≤ λ⊗Φ(...). Absolute continuity does not imply an ordering of measures; a probability density may exceed 1 (e.g., Uniform[0,1/2] has density 2). Thus the displayed inequality is false and the proof as written does not establish that the proposal transition preserves absolute continuity. This matters because the stationarity claim rests on the existence of the proposal density qγ. The theorem is nevertheless correct: the proposal is a Gaussian convolution, so for any measurable drift β, qγ(x,y) = (4πγ)^{-d/2} exp(-||y - x + γβ(x)||^2/(4γ)) is a density for each x, and ρeQγ has density ∫ρ(dx)qγ(x,y). Hence Γ = R+ and the C^2-a.e. / local-diffeomorphism apparatus is unnecessary. The paper should replace the flawed argument with this one-line Gaussian-convolution argument and remove the restrictive C^2 assumption from Theorem 2.13.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","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|.","tokens_in":15622,"tokens_out":10812,"duration_ms":123247,"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":[{"comment":"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.","section":"Section 2.2, proof of Theorem 2.13"},{"comment":"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.","section":"Theorem 2.13 statement; Section 1.2, eq. (1.8)"},{"comment":"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.","section":"Abstract, keywords; Section 3.1.2"},{"comment":"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.","section":"Section 1.2, eq. (1.8); proof of Theorem 2.13"}],"minor_comments":[{"comment":"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.","section":"Throughout"},{"comment":"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}}.","section":"Section 3.1.3"},{"comment":"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.","section":"Figure 4 caption"},{"comment":"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.'","section":"Section 3.1.1"},{"comment":"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.","section":"Section 2.2, proof of Theorem 2.13"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses an interesting and timely extension of MALA, but the central stationarity theorem has a genuine proof gap that is nonetheless easily fixable by a direct Gaussian-density argument. The main revisions are: correct the proof of Theorem 2.13, state the path-differentiability assumption explicitly, and align the 'exponential convergence/ergodicity' wording with the actual results, which are only stationarity plus numerical evidence. I do not see an irreparable flaw, so major revision rather than rejection seems appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nMASLA is a straightforward but sensible idea: run MALA's Metropolis correction with a subgradient drawn from a conservative field. As far as the cited literature goes, that specific combination is new — prior work either smooths the nonsmooth term or skips the accept-reject step. The paper's main theorem, stationarity and reversibility of the chain for almost every step size, is true. I checked the proof after the stress-test note flagged it, and the note is right: the displayed inequality ρeQγ(A) ≤ λ⊗Φ(...) does not follow from absolute continuity, since densities can exceed 1. But that is a proof bug, not a result bug. Because the proposal is a Gaussian convolution, the transition density qγ(x,y) exists for any measurable drift β(x); no C^2-everywhere-a.e. or local-diffeomorphism machinery is needed. So Theorem 2.13 holds, and the assumption that U is twice differentiable almost everywhere is unnecessarily restrictive.\n\nWhat the paper does well: it locates the right background (Bolte–Pauwels conservative fields, Habring et al.'s subgradient Langevin methods), and it honestly reports that MASLA matches Grad-sub and Prox-sub in the TV-L2 benchmark rather than beating them. That is a useful existence proof: the MH correction can be added to subgradient proposals without hurting performance. The numerical comparison with the unadjusted version (USLA) on |x^2-1| also illustrates the value of the accept-reject step.\n\nWhere it is soft: there is no convergence or ergodicity analysis. The keyword 'Exponential convergence' appears in the abstract, but the only support is empirical plots. That overstates the theory. The experiments are also under-reported: no code, no error bars, no clear description of how many chains or seeds went into Figures 4–6, and there are typos (e.g., 'discartion'). For a method paper that's fixable, but it needs to be addressed.\n\nOverall: the central sampler is plausible and the stationarity theorem stands, though the proof needs a one-line replacement. The contribution is modest — this is an incremental but legitimate extension, not a breakthrough. A serious referee should see it; the author should be asked to fix the proof, add a proper convergence discussion (even a conjecture or a reviewer-accessible condition), and make the numerics reproducible.\n\nBring it to reading group if you want a clean example of 'MH makes any proposal work'; otherwise skip. I would not cite it unless I needed the specific MASLA formulation.","headline":"A correct but modest extension of MALA to nonsmooth targets; the stationarity proof has a fixable flaw, and the paper overstates convergence support.","tokens_in":16225,"tokens_out":2355,"would_cite":false,"duration_ms":24632,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60J22","62F15","65C05","49J52"],"pacs":[],"model":"deepseek-v4-flash","headline":"MASLA, a subgradient-based Metropolis-adjusted Langevin sampler, is stationary and reversible for nonsmooth nonconvex targets under mild regularity.","keywords":["Metropolis-adjusted Langevin algorithm","conservative fields","subgradient MCMC","nonsmooth potentials","nonconvex sampling","path differentiability","reversibility","Markov chain Monte Carlo"],"falsifier":"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.","tokens_in":1886,"feed_emoji":"🎲","tokens_out":7540,"duration_ms":195540,"temperature":0.7,"pith_summary":"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.","feed_headline":"Subgradient sampler extends MALA to nonsmooth targets","feed_subtitle":"Metropolis correction plus subgradient proposals keeps the target stationary for almost every step size.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the conservative-field calculus, chain rule, and path-differentiability framework that MASLA's proposal uses.","marker":"Bolte and Pauwels (2021)"},{"why":"Establishes existence of strong solutions to the continuous-time Langevin differential inclusion whose Euler discretization MASLA corrects.","marker":"Difonzo et al. (2022)"},{"why":"Motivates the Metropolis adjustment by showing unadjusted Langevin discretizations can be transient, and supplies the density/detailed-balance proof template.","marker":"Roberts and Tweedie (1996)"},{"why":"Provides the proximal/subgradient Langevin baselines and the TV-L2 experimental setup used for comparison.","marker":"Habring et al. (2024b)"},{"why":"Provides the proximal MALA baseline that MASLA is compared against in total variation distance.","marker":"Pereyra (2016)"},{"why":"Provides the Moreau-Yosida ULA baseline and the smoothing approach that MASLA avoids by using subgradients directly.","marker":"Durmus et al. (2022)"}],"fun_headline_variants":["Metropolis correction subgradient Langevin for nonsmooth targets","Subgradient proposals plus Metropolis step make MALA nonsmooth-safe","Nonsmooth, nonconvex targets? MASLA samples them exactly","Metropolis-adjusted subgradient Langevin: valid for a.e. step size","MASLA: subgradient + Metropolis beats differentiability limits"],"cache_read_input_tokens":18304,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Metropolis correction subgradient Langevin for nonsmooth targets","Subgradient proposals plus Metropolis step make MALA nonsmooth-safe","Nonsmooth, nonconvex targets? MASLA samples them exactly","Metropolis-adjusted subgradient Langevin: valid for a.e. step size","MASLA: subgradient + Metropolis beats differentiability limits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000843,"raw_usage":{"total_tokens":3633,"prompt_tokens":868,"completion_tokens":2765,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":484,"completion_tokens_details":{"reasoning_tokens":2670}},"tokens_in":484,"tokens_out":2765,"duration_ms":23026,"temperature":1.0,"reasoning_tokens":2670,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:51:51.772771+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"and Pauwels, E","cited_arxiv_id":null,"evidence_quote":"Supplies the conservative-field calculus, chain rule, and path-differentiability framework that MASLA's proposal uses."},{"cited_title":"Stochastic Langevin Differential Inclusions with Applications to Machine Learning","cited_arxiv_id":"2206.11533","evidence_quote":"Establishes existence of strong solutions to the continuous-time Langevin differential inclusion whose Euler discretization MASLA corrects."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Motivates the Metropolis adjustment by showing unadjusted Langevin discretizations can be transient, and supplies the density/detailed-balance proof template."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the proximal MALA baseline that MASLA is compared against in total variation distance."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Moreau-Yosida ULA baseline and the smoothing approach that MASLA avoids by using subgradients directly."}],"review_version":1}