{"id":"341c28b9-7c08-4457-b4c5-615ae53c2261","arxiv_id":"2508.17545","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A higher-order Langevin Monte Carlo algorithm family is proposed with Wasserstein mixing time improving in the dynamics order, under a strong Condition H2 that the abstract does not disclose.","lead":"The paper designs order-P Langevin Monte Carlo samplers for any P≥3 and claims a mixing time that improves as P grows. The proof supports a weaker rate than the abstract states, and the improved rate relies on a very strong smoothness condition that is mostly left out of the abstract.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2.19 and the abstract advertise an ε^{-1/(2R)} mixing-time exponent, but the paper's own Proposition 2.26, when solved for η, yields ε^{-2/R}; Theorem 2.19's proof is also omitted, so the advertised rate is not established.","rationale":"The reader's CONDITIONAL verdict is appropriate. The strongest concern is not a subtle hidden assumption but the central display: the paper's own Proposition 2.26, when solved, gives a different ε-exponent than the abstract and Theorem 2.19. This is an internal inconsistency, not a matter of external consensus; it directly affects the claimed improvement. Theorem 2.9's P = 4 formula is explicit and confirms the 2/7 exponent, so the 1/(2R) in Theorem 2.19 is either a typo or an unsupported assertion. The empty proof of Theorem 2.19 means that even the corrected general-P statement lacks a written argument, though the structure suggests it follows from Proposition 2.26 by the same argument as Theorem 2.9. Condition H2 is a genuine limitation: it contains η, is stronger than MMW+21's Assumption 2, and is omitted from the abstract. However, the authors state H2 in the theorem and discuss it in Remarks 2.4-2.8, so it is a disclosed restriction rather than a hidden flaw. The numerical experiments use quadratic or low-degree Taylor approximations, so they do not validate the general-smoothness claim. Overall, a revised version that corrects the exponent, states H2 in the abstract, and actually writes out the proof of Theorem 2.19 would make the paper's corrected claim defensible. Hence the reader's CONDITIONAL verdict remains appropriate; no change is needed from the reader's assessment.","tokens_in":62626,"tokens_out":5050,"duration_ms":49798,"concrete_test":"Re-derive Theorem 2.19 from Proposition 2.26: set the number of iterations as k ≈ (1/(hη)) log(2 eC4 E[|Z−x(0)|^2]/ε^2) and choose the stepsize that makes eC3 d η^R = ε^2/2, namely η = (ε^2/(2 eC3 d))^{1/R}. Substitution gives a mixing time proportional to d^{1/R}/ε^{2/R}, not d^{1/R}/ε^{1/(2R)}. A second cross-checkable instance is Theorem 2.9's explicit d^{1/7}/ε^{2/7}, which already contradicts the abstract's ε^{-1/(14)} for P = 4.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Proposition 2.26 bounds the one-step discretization error by eC3 d η^R, where R = 4 for P = 3 and R = 2P−1 for P ≥ 4. Theorem 2.19 then sets η0 = (ε^2/(2 eC1 d))^{1/R} and claims k* = (2 eC3)^{1/R}/h · d^{1/R}/ε^{1/(2R)}. But solving eC3 d η^R ≤ ε^2/2 for η gives η ≤ (ε^2/(2 eC3 d))^{1/R}, and the contraction term requires k ≈ 1/(hη), hence k* ∼ d^{1/R}/ε^{2/R}. Theorem 2.9 itself states the P = 4 rate as d^{1/7}/ε^{2/7}, not d^{1/7}/ε^{1/14}. The abstract and Theorem 2.19 therefore state a rate that the proof does not deliver; the 2/R exponent is the correct consequence of the displayed equations. Compounding this, the proof of Theorem 2.19 in the Appendix is literally absent (\"Proof of Theorem 2.19. □\"), so the general-P claim is not actually proven in the text. Condition H2, which contains η on the right-hand side and is acknowledged in Remark 2.7 to be much stronger than Assumption 2 of MMW+21, is also omitted from the abstract, so the advertised \"convex-smooth\" guarantee is misleading; nevertheless the exponent error is the load-bearing defect.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes P-th order Langevin Monte Carlo algorithms (P >= 3) based on splitting discretizations of generalized Langevin dynamics, with the gradient inside each stage replaced by a Taylor polynomial. Under an m-strongly convex, L-smooth potential (Condition H1) and an additional high-order smoothness/growth condition (Condition H2), it claims 2-Wasserstein mixing time O(d^{1/R}/epsilon^{1/(2R)}) with R=4 for P=3 and R=2P-1 for P>=4, thereby improving on the third-order result of MMW+21. The proof borrows a continuous-time contraction estimate from Monmarché, bounds the one-step discretization error against the continuous dynamics, and then solves for the step size and number of iterations. Numerical experiments on Bayesian linear and logistic regression compare third- and fourth-order variants.","tokens_in":62867,"tokens_out":4778,"duration_ms":49289,"significance":"If the advertised rate were established, the paper would make a useful incremental contribution: it gives a concrete, implementable family of high-order LMC samplers and a plausible proof route via splitting plus Taylor approximation, with a dimension-independent contraction estimate imported in a transparent way. The authors are explicit that Condition H2 is stronger than the corresponding assumption in MMW+21, and they provide worked formulas for the Gaussian updates in the fourth-order case. However, the central advertised scaling is not the scaling proved in the displayed equations: the proof yields epsilon^{-2/R}, not epsilon^{-1/(2R)}. Since the main quantitative claim of the abstract and Theorem 2.19 is the load-bearing contribution, the paper currently overstates its own result. The contribution remains valuable once the rate is corrected to epsilon^{-2/R} and the missing proofs are supplied.","major_comments":[{"comment":"The advertised mixing time O(d^{1/R}/epsilon^{1/(2R)}) is not what the proof gives. Proposition 2.26 bounds the discretization error by eC3 d eta^R (up to the contraction term). Solving eC3 d eta^R <= epsilon^2/2 for eta yields eta = (epsilon^2/(2 eC3 d))^{1/R}, and the contraction term e^{-(k+1)h eta} <= epsilon^2/2 requires k of order 1/(h eta), so k* is of order d^{1/R}/epsilon^{2/R}. This is exactly the exponent stated in Theorem 2.9 for P=4, namely epsilon^{2/7}, so the 1/(2R) exponent in the abstract and in Theorem 2.19 is internally inconsistent with the paper's own equations and should be corrected to 2/R.","section":"Abstract and Theorem 2.19"},{"comment":"The proof of the general-P theorem is absent: the appendix contains literally 'Proof of Theorem 2.19. □', and the main-text proof only says the argument is the same as for Theorem 2.9. Since the general-P discretization error in Proposition 2.26 is the load-bearing ingredient and its proof depends on the skipped details of Lemma 2.25, the theorem is not proven as printed. A complete derivation of the claimed mixing time from Proposition 2.26 must be supplied.","section":"Appendix C, 'Proof of Theorem 2.19'"},{"comment":"The proof of Lemma C.4 is omitted with the statement that it is very similar to the proof of Lemma B.4. This lemma provides the moment bound E||x(k+1)||^{2alpha} <= (eC1)^alpha (d+2alpha)^alpha that is used in Lemma 2.25 and hence in Proposition 2.26. Because the stage-wise difference exponents in Lemma 2.25 rely on the combination of this moment bound with Condition H2, the omission concerns a load-bearing ingredient rather than a routine verification. A full proof should be included in a revised version.","section":"Lemma C.4"},{"comment":"Condition H2 is load-bearing for the stated improvement, but the abstract and the contribution bullet in Section 1 advertise the result as holding for convex-smooth potentials without conveying the assumption's strength. The condition contains eta on the right-hand side and, as Remark 2.7 acknowledges, is much stronger than Assumption 2 of MMW+21. The advertised 'convex-smooth' guarantee is therefore misleading; the condition should be stated explicitly wherever the main rate is announced.","section":"Condition H2 and abstract"}],"minor_comments":[{"comment":"The statement displays the discretization bound as C3 d eta^8, while the proof concludes with C3 d eta^7 and Theorem 2.9 solves C3 d eta^7 <= epsilon^2/2. These should be harmonized.","section":"Proposition 2.18"},{"comment":"The phrase 'it holds when and P>=4' is garbled and should read 'for P>=4' (and similarly a separate statement for P=3).","section":"Proposition 2.26"},{"comment":"The step-size definition uses eta0 := (epsilon^2/(2 eC1 d))^{1/R} with eC1 from Lemma C.4, while the discretization error that is solved for uses eC3 from Proposition 2.26; the notation should be aligned so that the displayed eta0 matches the constant in the equation actually solved.","section":"Theorem 2.19"},{"comment":"The right-hand side writes v^{stj-1}_n(t) where the index should presumably be P-j; the current notation is confusing and should be corrected.","section":"Lemma 2.25, item (b)"}],"recommendation":"major_revision","confidential_remarks":"The paper has a real algorithmic idea and the corrected rate d^{1/R}/epsilon^{2/R}, which is what Theorem 2.9 already states for P=4, still improves on MMW+21's d^{1/4}/epsilon^{1/2} for moderate P. The central problem is that the headline is wrong, not that the proof is nonsense; this is fixable by restating the abstract and Theorem 2.19, and by supplying the missing proofs of Theorem 2.19 and Lemma C.4. I would not reject, but the revision must be substantive rather than cosmetic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead the paper. Two things up front. The algorithm family is new: for any P>=4, a P-stage splitting discretization of P-th order Langevin dynamics with Gaussian conditional updates, and the observation that P-1 is the maximum number of stages is a useful design principle. The continuous-time contraction is taken from Monmarche, so the real work is the discretization analysis. The P=3 case is close to MMW+21 with Taylor instead of Lagrange interpolation.\n\nThe paper also has a load-bearing mistake in the advertised rate. The abstract and Theorem 2.19 say mixing time O(d^{1/R}/eps^{1/(2R)}). But Proposition 2.26 bounds the one-step error by eC3 d eta^R, and setting that equal to eps^2/2 gives eta ~ (eps^2/d)^{1/R}, hence k* ~ d^{1/R}/eps^{2/R}. Theorem 2.9 itself writes d^{1/7}/eps^{2/7} for P=4. So the correct exponent is 2/R, not 1/(2R). The corrected rate is still an improvement over MMW+21's d^{1/4}/eps^{1/2}, so the core claim survives, but the paper as written states a rate it does not prove.\n\nSecond, the proof of Theorem 2.19 is absent: the appendix contains 'Proof of Theorem 2.19. □' and nothing else. The argument would be identical to Theorem 2.9, so I read this as an omission rather than a fatal gap, but it needs to be written out.\n\nThird, the heavy assumption. Condition H2 is much stronger than MMW+21's Assumption 2, and the paper says so in Remark 2.7. It contains eta on the right-hand side, so as the target accuracy shrinks, it effectively restricts the guarantee to polynomial potentials (or functions satisfying a growth condition like (9)). The abstract says 'log-concave and smooth densities' without mentioning H2; that is misleading.\n\nOn the positive side, the proof architecture is plausible: contraction from Mon23 plus a standard Wasserstein perturbation argument. The stage-by-stage difference bounds in Lemmas 2.24 and 2.25 are elaborate and, from my check, internally consistent. The explicit P=4 formulas are a lot of work and look correct. Some appendix proofs are sketched rather than fully written (Lemma C.4, the induction in Lemma 2.25), which is a quality issue but not a red flag. The numerics are illustrative and use low-degree Taylor approximation outside the theorem's scope; the authors admit this.\n\nWho is this for: people working on high-order Langevin samplers and hypocoercive discretizations. It deserves a serious referee. I would send it to review, asking for the exponent fix, a written proof of Theorem 2.19, and an abstract that states H2. The corrected contribution is worth having in the literature, if not in its present form.","headline":"The construction is real, but the advertised epsilon exponent is not the one proved; with the corrected 2/R rate and the strong H2 condition disclosed, this is a credible within-subfield extension that needs revision before acceptance.","tokens_in":63551,"tokens_out":3617,"would_cite":false,"duration_ms":35521,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes $P$-th order Langevin Monte Carlo algorithms and claims mixing times of order $O(d^{1/R}/\\epsilon^{1/(2R)})$ for $R=4$ ($P=3$) and $R=2P-1$ ($P\\ge 4$).","keywords":["Langevin Monte Carlo","high-order Langevin dynamics","log-concave sampling","Wasserstein convergence","mixing time","splitting scheme","polynomial gradient approximation","Markov chain Monte Carlo"],"falsifier":"Check the exponent by solving the paper's constraint $C d \\eta^R \\le \\epsilon^2/2$ for $\\eta$ and substituting into the mixing-time factor $1/(h\\eta)$; the $\\epsilon$-exponent this returns can be compared with the theorem's stated $\\epsilon^{-1/(2R)}$. Numerically, run the fourth-order algorithm on a quadratic potential with known covariance, record the largest stepsize reaching a fixed $W_2$ error, and compare that scaling with the theorem's formula.","tokens_in":62266,"feed_emoji":"🎲","tokens_out":16027,"duration_ms":137669,"temperature":0.7,"pith_summary":"This paper proposes a family of sampling algorithms based on discretizing $P$-th order Langevin dynamics for any $P\\ge 3$, and argues that the mixing time improves as $P$ grows. Each step of the algorithm splits the dynamics into $P-1$ stages: the nonlinear coordinate is updated from a high-degree Taylor approximation of the gradient, while the remaining coordinates are integrated exactly as a linear Ornstein-Uhlenbeck process, so the next iterate is conditionally Gaussian. For strongly convex, $L$-smooth potentials satisfying an extra high-order derivative-growth condition, the claimed 2-Wasserstein mixing time is $O(d^{1/R}/\\epsilon^{1/(2R)})$ with $R=4$ for $P=3$ and $R=2P-1$ for $P\\ge 4$. If correct, this gives an explicit positive answer to the conjecture that higher-order Langevin dynamics can reduce both the dimension and accuracy dependence of log-concave sampling.","feed_headline":"P-th order Langevin sampling claims faster mixing as P grows","feed_subtitle":"A stage-splitting construction gives explicit Gaussian updates, improving mixing time in both dimension and accuracy.","key_machinery":"The load-bearing object is the $P-1$-stage splitting discretization of the $P$-th order Langevin dynamics. In each stage, the coordinate that carries the nonlinear potential, $v_1$, is advanced with a polynomial approximation $g^{\\mathrm{st}_j}(t)=P_{\\alpha-1}(\\theta^{\\mathrm{st}_{j-1}}(t))$ of the gradient, while the remaining coordinates evolve as a linear multivariate Ornstein-Uhlenbeck process that can be integrated exactly; the whole scheme is built so that $\\mathbb{E}[x^{(k+1)}\\mid x^{(k)}]$ is multivariate normal. The argument also uses a symmetric positive-definite matrix $M$ from the continuous-time contraction theory, which supplies a dimension-free dissipation rate $\\rho$ and lets the discretization error be tracked in the $M$-norm. Condition H2 is the matching condition that makes the Taylor remainder contribute at the same order $\\eta^R$ as the splitting error; without it the polynomial approximation would dominate and the rate would not improve with $P$.","core_discovery":"Under Condition H1 ($U$ is $m$-strongly convex and $L$-smooth) and Condition H2 (the scaled operator norms $L_\\alpha/\\alpha!$ of the high-order derivatives are controlled relative to $d$ and the stepsize), the paper's $P$-th order Langevin Monte Carlo algorithm is claimed to have 2-Wasserstein mixing time $O(d^{1/R}/\\epsilon^{1/(2R)})$ with $R=4\\cdot 1_{\\{P=3\\}}+(2P-1)\\cdot 1_{\\{P\\ge 4\\}}$. The algorithm is a $P-1$-stage splitting discretization: at each stage the velocity component that feels the potential is updated using a degree $\\alpha-1$ multivariate Taylor polynomial of $\\nabla U$, the remaining components are integrated as an exactly solvable Ornstein-Uhlenbeck process, and the next iterate is drawn from the resulting multivariate normal distribution. The proof bounds the one-step discretization error by $C d \\eta^R$ plus an exponentially decaying initialization term, using a dimension-free contraction matrix for the continuous dynamics; Condition H2 is exactly the statement that the polynomial approximation error is no larger than the splitting error.","pith_inferences":["The same Gaussian-transition construction should extend to stochastic-gradient settings, with an extra gradient-noise term added to the discretization error; the paper does not analyze this variant.","The $P-1$-stage limit suggests a tradeoff: more stages raise the splitting order but require stronger smoothness, so for non-polynomial potentials the optimal order may be finite rather than the $P\\to\\infty$ improvement suggested by the rate formula.","A non-polynomial implementation with higher-degree Taylor or Chebyshev approximations could be tested against the fourth-order algorithm's observed gains, isolating how much of the improvement comes from the splitting order versus the polynomial degree.","The dimension-free contraction constants suggest that in practice the Condition H2 constant $c$ and the choice of friction $\\gamma$ will matter as much as the asymptotic $d^{1/R}$ term."],"forward_implications":["For $P=4$, the paper claims a mixing time of $O(d^{1/7}/\\epsilon^{1/14})$, improving on the third-order rate $O(d^{1/4}/\\epsilon^{1/2})$ in both dimension and accuracy.","Each iteration is a single Gaussian draw with closed-form mean and covariance when the potential is quadratic, so the per-step cost is set by sampling a $Pd$-dimensional normal and forming the polynomial coefficients.","For non-polynomial smooth losses such as logistic regression, the required Taylor coefficients are approximated to fixed degree, which is what the numerical experiments use.","If the central claim holds, it settles the conjecture that high-order Langevin discretizations reduce the dependence on dimension and accuracy, at least for potentials satisfying Condition H2.","As $P$ increases, $R=2P-1$ grows linearly, so the claimed dimension and accuracy exponents improve; the practical bottleneck moves to computing higher-degree Taylor coefficients of $\\nabla U$."],"supporting_citations":[{"why":"Supplies the continuous-time contraction inequality and the dimension-free matrix $M$ on which the discrete mixing-time proof rests.","marker":"[Mon23]"},{"why":"Provides the third-order Langevin Monte Carlo algorithm, its rates, and the conjecture about high-order acceleration that this paper targets.","marker":"[MMW+21]"},{"why":"Contributes the underdamped Langevin splitting methodology and the second-order mixing-time comparison point in the table.","marker":"[CCBJ18]"},{"why":"Gives the randomized-midpoint mixing rate listed in the comparison table for convex-smooth potentials.","marker":"[SL19]"},{"why":"Provides the multivariate Taylor theorem with remainder used to construct the polynomial approximations of the gradient.","marker":"[Car71]"},{"why":"Supplies the explicit coefficient formula for multivariate Taylor expansion used to define the approximating polynomials.","marker":"[FA89]"},{"why":"Offers the derivative-growth condition, expressed through the gamma function, that implies Condition H2 for non-polynomial potentials.","marker":"[WWJ16]"}],"fun_headline_variants":["Higher-order Langevin cuts sampling cost in dimension and accuracy","P-th order Langevin: faster mixing as order grows","Langevin at order P tightens dimension dependence in mixing","Higher-order LMC improves tradeoff between dimension and error","P-th order splitting yields explicit Gaussian updates and faster mixing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole rate rests on Condition H2: the potential's very high derivatives, after factorial scaling, must stay small enough relative to the dimension and the stepsize that a polynomial approximation of the gradient is no worse than the splitting error; for non-polynomial smooth potentials this becomes harder to satisfy as the target accuracy shrinks.","fun_headline_variants_meta":{"raw":{"variants":["Higher-order Langevin cuts sampling cost in dimension and accuracy","P-th order Langevin: faster mixing as order grows","Langevin at order P tightens dimension dependence in mixing","Higher-order LMC improves tradeoff between dimension and error","P-th order splitting yields explicit Gaussian updates and faster mixing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000227,"raw_usage":{"total_tokens":1483,"prompt_tokens":967,"completion_tokens":516,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":583,"completion_tokens_details":{"reasoning_tokens":434}},"tokens_in":583,"tokens_out":516,"duration_ms":5475,"temperature":1.0,"reasoning_tokens":434,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:06:04.418431+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Check the exponent by solving the paper's constraint $C d \\eta^R \\le \\epsilon^2/2$ for $\\eta$ and substituting into the mixing-time factor $1/(h\\eta)$; the $\\epsilon$-exponent this returns can be compared with the theorem's stated $\\epsilon^{-1/(2R)}$. Numerically, run the fourth-order algorithm on a quadratic potential with known covariance, record the largest stepsize reaching a fixed $W_2$ error, and compare that scaling with the theorem's formula.","supporting_citations":[],"review_version":2}