- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
Two algorithms, Proximal MUON Langevin dynamic (PMLD) and Annealed PMLD (APMLD) are introduced to incorporate the geometry induced by the MUON optimizer into the Langevin dynamics for improving Bayesian learning tasks. For the first algorithm, PMLD, the authors design a version of the Unadjusted Langevin algorithm (ULA) that targets a "modified" posterior measure involving the original objective and the Moreau-Yosida envelope of the indicator function of a spectral-norm ball. In this way, the spectral constraint is encoded through a smooth prior. The aim of the second algorithm, APMLD, is to propose "a more informative way" to initialize the sampling to improve the finite-time behaviour. In particular the algorithm runs MUON on the smoothed potential during an initial phase, then selects the iterate associated with the smallest (nuclear) norm of the gradient of the smoothed potential, applies a Gaussian perturbation, and then run standard ULA to the chosen target. A convergence analysis is provided under strong convexity and smoothness assumptions separating the equilibrium spectral localization from the (finite-time) sampling error of ULA. Numerical experiments on Bayesian classifiers show the advantage of PMLD in terms of robustness comparing with ULA on Bayesian diffusion models and Bayesian classifiers. Moreover, APMLD shows good early posterior-predictive performance.
Critical strengths
- To the best of my knowledge, this work is one of the first attempts in the literature to study the MUON's spectral geometry within the framework of the Langevin dynamics, which is a highly non-trivial and important research problem. I would add another perspective of the importance of this work beyond the motivation for Bayesian learning stated in the manuscript by the authors. Stochastic MUON with fixed mini-batches may fail to converge to a global minimizer even for smooth and strongly convex objectives, see, e.g., [1]. Moreover, non-convergence examples for convex Lipschitz functions have been appeared recently in the literature, see, e.g., [2]. Therefore, it is crucial to investigate how Muon's spectral geometry can be incorporated into stochastic frameworks whose long-time behaviour can be rigorously characterized with stochastic analysis tools. Such direction may provide alternative approaches to studying MUON-inspired algorithms and overcome some of the limitations of the current optimization guarantees in the literature. While the proposed PMLD algorithm is designed for Bayesian learning tasks (and it is in essence a standard ULA applied to a modified target) and for this reason, it does not directly address the issue of the convergence of stochastic MUON, I believe that the idea of the authors behind PMLD could provide an important stepping stone for further studies of MUON-"inspired" Langevin algorithms, which will be of benefit not only to the sampling community, but in particular to the optimization community.
References:
[1] On MUON optimization: From non-convergence to an error analysis with Polar Express and the Newton-Schulz polynomial from implementations, Thang Do, Steffen Dereich, and Arnulf Jentzen
[2] Muon Does Not Converge on Convex Lipschitz Functions, Tetiana Parshakova, Ahmed Khaled, Michael Crawshaw, Guillaume Garrigos, Robert M. Gower
- The authors manage to show that, outside the prescribed spectral region, the posterior associated with the PMLD target (composed of the original objective and the Moreau-Yosida envelope) has exponential decay in the squared spectral violation. This is important because it allows to quantify the contribution of the spectral prior and then later in the finite-time analysis separate the equilibrium spectral localization (MUON part) from the finite-time sampling error (ULA part). This contribution may be useful for the subsequent analysis of convergence of stochastic MUON suggested in point 1 above.
Critical weaknesses
-
(Critical Weakness): Limited discussion of the comparison with prior work on ULA/MALA/proximal gradient algorithms in the convergence analysis, including the comparison provided in Table 1 (page 2). More clarification is needed for positioning the proposed algorithms. More discussion about the "representative rate" is needed, particularly regarding the contribution of the modified target in PMLD and the differences between the target distributions considered in prior work. See question 1 below.
-
The convergence analysis for ULA is provided under very strong assumptions (strong convexity and smoothness). These assumptions are presented as necessary at page 4: "while the finite-time ULA guarantees additionally require the global convexity and smoothness assumptions stated below." The literature on the convergence analysis of ULA is quite well-established beyond strong convexity and smoothness, see, e.g., "From stability of Langevin diffusion to convergence of proximal MCMC for non-log-concave sampling" NeurIPS 2025, and "The Performance Of The Unadjusted Langevin Algorithm Without Smoothness Assumptions" TMLR, 2025 and references therein for the results in the nonconvex and nonsmooth settings. This is a limitation of the current convergence analysis, which does not cover the nonconvex settings under which the experiments are conducted. The authors state this limitation in the text, but still it remains unclear how the theoretical finite-time guarantees can explain the behavior of the experiments shown in the numerical experiments.
Questions for authors
- In Table 1 and in the text below the table, the comparison of the "representative rate" is very limited, particularly regarding the different target distributions (and assumptions) considered in prior works. Some of the parameters in Table 1 are not specified.
(a) For the RGO sampler result stated in the table, what other terms are contained in ? Can you explan why the dependence in the dimension in the PMLD and APMLD convergence analysis is worse comparing to the result of Liu and Chewi, (2026) for the RGO sampler (taking into account the different oracle assumptiosn)?
(b) What is the price in terms of the dimension and accuracy that we have to pay when using PMLD and APMLD instead of Proximal Langevin dynamics and Proximal Muon Langevin dynamics under Assumption 4.1? For PMLD how does the dependence on the parameter (Moreou-Yosida parameter) has an effect on this comparison?
(c) In addition, can you please clarify this sentence (Page 2, lines 97-98):
"Table 1 records representative guarantees together with their initialization requirements. " Where the initialization requirements are stated in Table 1?
-
In proximal Muon Langevin dynamics, how can we choose the spectral cap ? Do you need to be tuned? Is it problem specific as in Section 5.1.1? Is there a way that I can select it based on the values of or if the inverse temperature/temperature parameter is set to be different from one in ULA (i.e. or with ) can it be related to it? Or do we need to tune separately and ?
-
Section 5.2 "Bayesian Diffusion Models" compares the performance of PMLD/APMLD with SGLD rather than ULA. Can you please comment whether the convergence analysis for the PMLD can be extended to PMLD with the stochastic gradient version of PMLD? The theory of non-asymptotic convergence analysis of SGLD is quite now well-established for results beyond convexity and standard smoothness assumption.
-
(Suggestion) It would be helpful to summarize what are the main difficulties in proving convergence analysis for PMLD and APMLD comparing to ULA, as the proof in Appendix F seems to follow standard arguments for the ULA convergence analysis.
-
(Suggestions to improve clarity) It would be helpful to provide a more detailed derivation for the steps summarized in the proof of Lemma F.2 in lines 890-897 at page 17.
-
(Suggestions) I would suggest including in Proposition 4.3 the Big-O expression of the bounds for the discretization error and the initialization error and mixing terms provided in Theorem F.1. As it is stated right now, the finite-time sampling bound in Proposition 4.3 is standard for ULA and is not very informative about the specific properties of the PMLD and APMLD. Lemma F.2, Theorem F.1: It would be helpful to state and highlight the restriction on the step size provided in these two statements in the main text (Section 4) since depends on for PMLD algorithm.
5 comments
(2/2)
Q4. We agree that Appendix F follows standard ULA arguments. The method-specific contributions concern spectral target localization for PMLD and the conversion of MUON's optimization progress into quantitative initialization bounds for APMLD. Our common sampling analysis deliberately separates these contributions from the subsequent Langevin dynamics, rather than introducing a new ULA convergence mechanism.
Q5–Q6. Thank you for this suggestion. The paragraph provides a technical justification for applying the standard Girsanov-based discretization argument for Langevin Monte Carlo (see, e.g., Section 4.4 of Chewi's Log-Concave Sampling). The localization step ensures that the relative entropy inequality is valid under our finite-second-moment assumption, without requiring additional exponential integrability of the unstopped drift difference. We will expand the derivation by explicitly defining the stopping time, stating the Girsanov entropy identity for the stopped path laws, and justifying the passage to the unstopped laws.
The explicit discretization and initialization-dependent bounds, together with the step size restriction, are already established in Lemma F.2 and Theorem F.1. Section 4.3 also discusses the trade-off between spectral localization and the increased smoothness constant induced by smaller Moreau parameters (lines 340–344).
We agree, however, that incorporating these quantitative estimates directly into Proposition 4.3 would make the method-specific implications more transparent. We will therefore include the explicit discretization bound from Theorem F.1 and highlight the step size restriction in the main text, particularly its dependence on the Moreau parameter for PMLD. This will better illustrate how PMLD affects the target smoothness and discretization error, whereas APMLD affects the initialization-dependent terms.
Reply to Authors
Thank you very much for the detailed rebuttal.
Q2 in (1/2): I have small a follow-up question about this point. In practice, can you identify a rule for choosing and/or ? Have you observed any behaviour/pattern for choosing the values for these parameters, e.g. through a sensitivity analysis in Experiment 5.1.1? If not possible, did you get any insights into how to choose and from the derivation of Proposition 4.1?
Q1(b) in (1/2): OK, thank you for clarifying this. I suggest to report explicitly the value of in the Table 1 if possible to highlight this point about the trade-off to the readers since the aim is to position the proposed algorithms within the broader sampling literature (as point it out in the answer to Q1 (c)).
Q1(a), Q1(c), and Q3 in (1/2): OK.
Q4- Q6 in (2/2): I thank the authors for taking into account my suggestions.
(1/2)
We sincerely thank the reviewer for the detailed technical comments and helpful references! We agree that the convergence comparison and the presentation of our theoretical results can be substantially clarified.
Q1(a). Their Theorem 3.1 establishes a TV-accuracy complexity of O(κdlog4(1/ε)), where κ=max(1,β/α), and O suppresses additional logarithmic factors in d and κ. Their algorithm uses an augmented-space proximal sampler with Metropolis–Hastings corrections, access to the mode, and a restricted Gaussian oracle (RGO) for g. For our spectral-ball indicator, this oracle requires sampling a Gaussian restricted to the spectral-norm ball, which is different from evaluating its projection, and might not have implementation in a closed form . Their sharper dimension and accuracy dependence reflects both the oracle access and the sampling mechanism. Our methods instead retain explicit Euler–Langevin updates and their associated discretization error. We will clarify these differences rather than suggest a direct comparison of computational costs.
Q1(b). Under Assumption 4.1, PMLD retains the polynomial dimension and accuracy dependence O(κγ2d/ε2) of ULA, with κγ=(L+γ−1)/M. Holding initialization constants fixed, the condition-number factor increases relative to ordinary ULA by $\frac{\kappa_\gamma^2}{\kappa^2} =\left(1+\frac{1}{L\gamma}\right)^2.$ Thus, smaller γ strengthens spectral localization but worsens the conservative smoothness constant and restricts the sampling step size.
Q1(c). Thank you for identifying this oversight. We will add initialization and oracle requirements to Table 1. However, it is worth pointing out that our objective is not to improve upon the convergence rates of state-of-the-art samplers, but rather to provide a representative overview of related methods and position our proposed algorithms within the broader sampling literature.
Q2. These parameters have distinct roles: R specifies the desired spectral scale, whereas γ controls the softness of the constraint. In Section 5.1.1, we use the layer-specific choice Rℓ=0.5∣Wℓ,MAP∣2 and γ=10−4. Indeed, they can be tuned separately. Here our γ is the scale on prior only, while a temperature β or τ governs the whole posterior; they are compatible since the scales are multiplicative. For example, we can have
πγ,τ(∣W∣op>R+t)≤πτ(B(R))exp[−t2/(2γτ)],
where πτ∝e−U/τ is the tempered base distribution.
Q3. Since stochastic PMLD is SGLD applied to Uγ, existing SGLD analyses can be utilized when their assumptions hold. More concretely, under Assumption 4.1 and conditionally unbiased stochastic gradients with conditional variance bounded by σG2, synchronous coupling of full-gradient PMLD (Xk) and stochastic PMLD (Yk), starting from the same initialization, gives
\leq \sigma_G\sqrt{\frac{\eta(M+L_\gamma)}{2ML_\gamma}},$$ for $0<\eta\leq 2/(M+L_\gamma)$. This provides an explicit additional stochastic-gradient error in Wasserstein distance. We also appreciate the references on nonconvex and nonsmooth Langevin analysis. Global strong convexity and smoothness are assumptions of our particular analysis, not necessary conditions for ULA convergence generally. For example, Vempala and Wibisono (2019) establish ULA convergence under smoothness and a log-Sobolev inequality without global convexity. Since PMLD and the post-switch phase of APMLD reduce to ULA, such results provide alternative sampling guarantees under their respective assumptions. We will clarify this broader applicability while acknowledging that our current quantitative guarantees do not directly cover the experimental neural-network posteriors.