A hierarchical entropy method for the delocalization of bias in high-dimensional Langevin Monte Carlo

Daniel Lacker; Fuzhong Zhou

REVIEW 1 major objections 4 minor 1 cited by

A hierarchical entropy method for the delocalization of bias in high-dimensional Langevin Monte Carlo

T0 review · 1 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The paper proves that for sparse or weakly interacting targets, the bias of unadjusted Langevin Monte Carlo in any k-dimensional marginal is O(hk) in relative entropy, with constants independent of ambient dimension and no logarithmic facto

desk verdict Genuine improvement over Chen et al. 2024, clearly written, with a real scope caveat in the conditional Talagrand assumption that the authors themselves flag. read the letter →

arxiv 2509.08619 v1 pith:XVIIYCYF submitted 2025-09-10 stat.ML cs.LGmath.PR

Daniel Lacker , Fuzhong Zhou This is my paper

classification stat.MLcs.LGmath.PR MSC 60J6065C05

keywords delocalizationofbiasLangevinMonteCarlorelativeentropymarginalsparseinteractionsweaklogarithmicSobolevinequalityhierarchicalmethod

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

The paper proves that the bias of the unadjusted Langevin algorithm is delocalized: for targets whose coordinates interact sparsely (or weakly), the error in any k-dimensional marginal scales like the step size h times k, with no dependence on the ambient dimension n and no logarithmic factor. This improves the recent delocalization bound of [4], which was O(hk log n) in squared Wasserstein distance, and it replaces the standard O(nh) joint bias that forces step sizes to shrink like 1/n. The new bound is stated in relative entropy, a stronger distance than Wasserstein, and it holds under a logarithmic Sobolev inequality plus a conditional transport inequality, rather than requiring strong log-concavity. The proof tracks the relative entropies of all marginals simultaneously: the set-indexed vector of marginal entropies obeys a hierarchical system of differential inequalities whose linear part is the generator of a Markov process on subsets of coordinates.

What carries the argument

The key object is the set-indexed vector of marginal relative entropies H_t(u) = H(ρ_t^u | π^u), u ⊆ [n]. Proposition 2.2 gives each entry dH_t(u)/dt ≤ A_1^t(u) + A_2^t(u) − α H_t(u). The hierarchy term A_2^t(u) is bounded, via the conditional Talagrand inequality, by λ(H_t(N(u)) − H_t(u)) with λ = γβ²/(αε); the linear part is the rate matrix of a Markov chain on subsets that jumps from u to N(u) at rate λ, so its semigroup is e^{tA}F(u) = E[F(N_Λ(t)(u))] for a Poisson process Λ of rate λ. Iterating the discrete-time inequality expresses H_{kh} as sums of this semigroup applied to initial entropies and to the size function |u|, and the neighborhood-growth assumptions reduce to moment bounds

What would settle it

Take a strongly log-concave pairwise-interaction target of the form (1.11) with row sums Σ_{j≠i} ||V''_{ij}||_∞ just above the threshold α²/(γM_0) required by assumption (B.3), on a complete graph, and compute H(π_h^{{i}} | π^{{i}}) as n grows with h slightly below h*. If it exceeds C h times a factor growing with n (or log n), then the weak-interaction smallness condition is quantitatively necessary; if it stays O(h), the threshold can be relaxed. In the sparse regime, the analogue is to test a target satisfying (LSI), (A.1), (A.2) whose graph neighborhoods grow at rate r ≥ 1 + α²/(γβ²) and s

Watch

Extended reading notes

Core claim

The paper's central claim (Theorem 1.1) is that, under a logarithmic Sobolev inequality, a Lipschitz gradient bound, and a conditional Talagrand inequality, if the interaction graph defined by the Hessian has polynomially growing neighborhoods, then H(π_h^u | π^u) ≤ C h |u| for every coordinate set u and step size h ≤ h*, with C and h* depending only on the structural constants, not on the dimension n; Talagrand's inequality then gives (α/2) W_2^2(π_h^u, π^u) ≤ C h |u|. Theorem 1.2 extends this O(hk) entropy bound to graphs with exponential neighborhood growth up to the critical rate r < 1 + α²/(γβ²), and Theorem 1.5 proves the analogous bound for potentials V = Σ V_u(x_u) with weak interact

Load-bearing premise

The paper needs each conditional target measure to satisfy a Talagrand-type transport inequality (assumption (A.2), or (B.2) in the weak-interaction setting) with a dimension-free constant; this is the only input used to bound the hierarchy term that couples different marginals, and it does not follow from the global logarithmic Sobolev inequality when the target is not log-concave.

Editorial extensions

If this is right

For a k-dimensional observable of a sparse- or weak-interaction target, one can use a step size of order 1/(C k) and still get O(hk) total bias, instead of 1/n, making LMC practicable in very high dimension.
Because the bound is in relative entropy, Talagrand's inequality immediately yields the same O(hk) order for squared Wasserstein distance between k-marginals, with constant α/2.
The assumptions are strictly weaker than strong log-concavity and are preserved under bounded perturbations of the potential, so the delocalization extends to non-log-concave targets.
In continuous time, the marginal entropy decays at rate 2α(1−ε) but with the initial entropy evaluated at the random neighborhood N_Λ(t/ε)(u); if initial entropies grow at most linearly with marginal size, the delocalized decay e^{-2αt}|u| follows.
The weak interaction class covers pairwise and higher-order interaction potentials with small interaction strength, giving a new regime beyond sparse-zeros in the Hessian.

Reading between the lines

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

The hierarchical set-indexed method is not specific to LMC: any discretized diffusion or MCMC scheme whose drift has a local dependency graph and whose conditional measures satisfy transport inequalities should exhibit a comparable O(h|u|) delocalization; testing this on Metropolis-adjusted or preconditioned schemes is a direct next step.
The critical growth threshold r < 1 + α²/(γβ²) in Theorem 1.2 has the flavor of a phase transition: beyond it, the Poisson-driven expansion of neighborhoods may outrun the entropy contraction, so a log n factor (or worse) should reappear. Building the missing counterexample would settle the sharpness question the paper leaves open.
In the weak-interaction result, the interaction strength R_1 only appears in the smallness condition (B.3), not in the final constant C or h*, which suggests the condition might be relaxed by a finer non-commutative analysis; a numerical scan of row sums near the threshold would indicate how tight the condition is.
The paper's Section 1.4 notes that average-case delocalization is automatic by subadditivity; the contribution is the worst-case upgrade. This suggests testing worst-case delocalization for observables that are not additively structured, e.g., smooth functions of a fixed small coordinate subset with coupling to far-away coordinates through the drift.

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.

Desk Editor's Note

read the letter

Dan, here's my read. This paper delivers what it promises: for sparse interaction graphs it upgrades [4]'s W2^2 bound to relative entropy with O(hk) and no log factor, relaxes strong log-concavity to LSI plus a conditional transport assumption, and adds a separate weak-interaction class. The hierarchical entropy machinery is adapted from their prior propagation-of-chaos work, and it's applied competently. I checked the main differential inequalities in Sections 2-4; the constants are explicit, and the Poisson-process/operator formalism is a clean way to handle the set-indexed recursion. The paper is honest about where the work lies.

The one load-bearing nonstandard input is (A.2), the conditional Talagrand inequality on Markov-blanket conditionals. The stress-test note gets this right: global LSI does not imply conditional Talagrand with uniform constants, so the advertised 'relaxation' of strong log-concavity is not as broad as the phrase suggests. What saves this from being a flaw is that the paper states (A.2) explicitly, isolates its single use in Remark 3.4, and notes it can be weakened further to a drift-dependent condition. So the theorem is correct conditional on (A.2); the scope is narrower than 'general LSI targets.' That should be said in the final version, maybe with an example of a non-log-concave target satisfying it other than bounded perturbations of strongly log-concave measures.

The weak-interaction Theorems 1.4 and 1.5 are a nice addition. Theorem 1.4 is basically a short coupling argument following [4], attributed appropriately. Theorem 1.5's structural assumption (1.12) is restrictive, but that's in the nature of the result. Minor issue: Section 1.5 on RMT-style delocalization is disconnected from the rest of the paper; could be trimmed.

Citation pattern looks fine: [4] is the benchmark, and [9,10,11] supply the methodology, not the theorem. No data or code, but this is a theory paper. I'd send this to a serious referee; it deserves careful checking of the hierarchy estimates. I would cite it if I work on LMC bias or propagation of chaos. Reading group: yes.

Referee Report

1 major / 4 minor

Summary. The paper studies the bias of the unadjusted Langevin algorithm in high dimension. For potentials with sparse interactions, under a logarithmic Sobolev inequality (LSI), a Lipschitz gradient condition, and a conditional Talagrand inequality (A.2), it proves that the relative entropy between low-dimensional marginals of the LMC stationary distribution and the target decays as H(π_h^u | π^u) ≤ C h |u|, with no dependence on ambient dimension and no logarithmic factor. This strengthens the recent delocalization result of Chen et al. in the sparse regime, moves from squared Wasserstein distance to the stronger relative-entropy metric, and relaxes strong log-concavity. The paper also proves an analogous result under exponential graph growth with a critical growth exponent, a companion continuous-time Langevin result, and a separate delocalization result for a new class of weak-interaction potentials. The proofs introduce a hierarchy of set-indexed marginal entropies, derive a recursive differential inequality, and analyze it through an auxiliary Markov process on subsets of coordinates.

Significance. If correct, this is a substantial and interesting strengthening of the delocalization phenomenon: the O(hk) bias bound for k-dimensional marginals is dimension-free in the ambient dimension, removes the logarithmic factor present in prior work, and holds in relative entropy, which is stronger than Wasserstein distance. The constants are explicit, the hierarchical method is transparent, and the weak-interaction class is a genuinely new regime. The main caveat is that assumption (A.2), the conditional Talagrand inequality, is not implied by the global LSI and is used in an essential way to close the hierarchy; the result is therefore conditional on a uniform transport condition on all Markov-blanket conditionals. This is a scope limitation, not an internal inconsistency, and the authors partly acknowledge it in Remark 3.4, but it should be more explicitly foregrounded in the introduction and abstract.

major comments (1)

[§1.1, §3.2 (Eq. (3.3)), Remark 3.4] The conditional Talagrand inequality (A.2) is load-bearing: it is used exactly once in the sparse proof, at Eq. (3.3), to pass from the drift discrepancy in A_2^t(u) to a relative-entropy difference via Lemma 3.3 and the chain rule. It is not a consequence of (LSI) plus (A.1) plus graph growth. For a global-LSI measure, conditional LSI/Talagrand constants need not be uniform and can deteriorate under conditioning on rare events. Thus the advertised 'relaxation' of strong log-concavity is narrower than a general LSI theory: the theorem covers, e.g., bounded perturbations of strongly log-concave measures via Holley–Stroock, but not arbitrary LSI targets. This is not a proof flaw, since (A.2) is stated explicitly, but the abstract and introduction should qualify the relaxation claim and add a discussion of the scope and possible failure of (A.2). The same caveat applies to the weak-interact

minor comments (4)

[§1.4, Eq. (1.14)] The parenthetical claim that the subadditivity inequality 'holds for relative entropy in place of W_2^2' is false. Counterexample: on {0,1}^2, let π be uniform on {(0,0),(1,1)} and let μ put mass q on (0,0) and 1−q on (1,1), with q≠1/2. Then H(μ|π)=D(q||1/2), and both one-dimensional marginal entropies equal the same D(q||1/2), so the average marginal entropy equals H(μ|π), which is strictly larger than (1/2)H(μ|π). The W_2^2 statement is correct, but the relative-entropy analogue should be removed or replaced by a correct statement.
[§5, Theorem 1.4 proof] The displayed condition 'since h≤α/β^2≤1/β' is stronger than the theorem's hypothesis h≤1/β. The argument only needs h≤1/β so that the diagonal entries of R_D, which lie in [1−βh,1−αh], are nonnegative. Please correct the sentence to avoid an apparent inconsistency with the theorem statement.
[§4.2, Eq. (4.9)] In the weak-interaction estimate, when w⊂u the set w\u is empty, and Lemma 3.3 as stated for nonempty Euclidean spaces is not directly applicable. The corresponding contribution to A_2^t(u) is identically zero because ∇_w V_w depends only on coordinates already in u; the proof should say this separately for full rigor.
[§5, Eq. (5.3)] The proof of Theorem 1.4 delegates the key calculation for E|Y_{(k+1)h}−\bar Y_{(k+1)h}|_∞^2 to [4, (A.6)–(A.7)]. Since this estimate feeds directly into the final constant, consider reproducing the short calculation or stating it as a lemma so the paper is more self-contained.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the proof reduces the desired entropy bound to explicitly stated assumptions (LSI), (A.1), (A.2), and graph growth; no fitted parameter or self-citation chain carries the argument.

full rationale

The paper's derivation chain is self-contained conditional on its explicit hypotheses. The central bound H(π^u_h | π^u) ≤ C h |u| is obtained from a hierarchy of differential inequalities for marginal entropies, using only the stated logarithmic Sobolev inequality, Lipschitz gradient assumption, and conditional Talagrand inequality. The assumption (A.2) is exactly what is needed to close the hierarchy term A_2^t(u); it is stated as a hypothesis, not derived from or identified with the conclusion. The paper even marks where it is used (Remark 3.4), and the same transparency appears for (B.2) in Remark 4.2. This is a scope caveat — (A.2) is not implied by global LSI in general — but it is not circularity. The self-citations [9,10,11] supply the hierarchical/Feynman-Kac proof device, not the delocalization theorem, and the results are compared against the independent benchmark of Chen et al. [4]. No fitted parameter is renamed as a prediction; the constants are explicit and the stationarity bound follows by sending k → ∞ from a dynamic estimate. The auxiliary Markov process on subsets is a proof device, and its contraction properties are consequences of the growth assumptions, not of the target bound. Therefore no load-bearing step reduces to its own input by construction.

Assumptions & free parameters 0 free parameters · 10 assumptions · 0 invented entities

No fitted parameters appear; all constants are explicit functions of the assumption constants. The axioms are either standard mathematical results or stated domain assumptions. No new physical or ontological entities are postulated; the Poisson process and subset-valued Markov process are proof devices.

assumptions (10)

standard math Itô calculus and Fokker-Planck equation for the piecewise-linear interpolation of LMC
Used in Proposition 2.1 to derive the marginal Fokker-Planck equation (2.4).
standard math Existence of jointly measurable conditional expectation [3, Proposition 5.1]
Justifies the existence of the drift b^u in Proposition 2.1.
standard math Log-Sobolev implies Poincaré and Talagrand inequalities [13]
Used repeatedly, e.g. (2.1), to bound quadratic Wasserstein distance by relative entropy.
standard math Vempala-Wibisono entropy derivative identity [15]
Starting point of Proposition 2.2 for differentiating marginal relative entropies.
standard math Poisson raw moment bound [1, Corollary 1]
Used in Section 3.4 to control E Λ^p(t) in the polynomial-growth case.
standard math Sub-Gaussian bound for ∇V under π [12, Theorem 2.2] or [2, Theorem 1.2]
Used in Section 5 to prove the improved entropy bound (5.1).
domain assumption (LSI) logarithmic Sobolev inequality with constant α
Assumed in Theorems 1.1-1.5 and used to obtain the -αH(u) decay term.
domain assumption (A.2)/(B.2) conditional Talagrand inequality for conditional measures of π
The only place used to bound the hierarchy term A_2; not implied by global LSI alone.
domain assumption Sparse graph growth bounds: polynomial (1.7) or exponential with threshold (1.9)
Controls growth of neighborhoods and the action of the operator N in Section 3.
domain assumption Weak interaction condition γ M0 R1 < α^2 (B.3)
Ensures the decay exponent τ is positive in Theorem 4.1.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A hierarchical entropy method for the delocalization of bias in high-dimensional Langevin Monte Carlo." pith.science (2026). https://pith.science/paper/XVIIYCYF

@misc{pith2026250908619,
  author       = {Pith},
  title        = {Pith review of: A hierarchical entropy method for the delocalization of bias in high-dimensional Langevin Monte Carlo},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XVIIYCYF}},
  note         = {Machine review of arXiv:2509.08619}
}

read the original abstract

The unadjusted Langevin algorithm is widely used for sampling from complex high-dimensional distributions. It is well known to be biased, with the bias typically scaling linearly with the dimension when measured in squared Wasserstein distance. However, the recent paper of Chen et al. (2024) identifies an intriguing new delocalization effect: For a class of distributions with sparse interactions, the bias between low-dimensional marginals scales only with the lower dimension, not the full dimension. In this work, we strengthen the results of Chen et al. (2024) in the sparse interaction regime by removing a logarithmic factor, measuring distance in relative entropy (a.k.a. KL-divergence), and relaxing the strong log-concavity assumption. In addition, we expand the scope of the delocalization phenomenon by showing that it holds for a class of distributions with weak interactions. Our proofs are based on a hierarchical analysis of the marginal relative entropies, inspired by the authors' recent work on propagation of chaos.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Delocalization of bias in unadjusted Hamiltonian Monte Carlo and underdamped Langevin
stat.CO 2026-07 conditional novelty 7.0 of 10

Unadjusted HMC and BAOAB Langevin exhibit delocalization of bias: W2,ℓ∞ bias scales as O(h√log d) under weak/sparse interactions, so O(√K) integration steps control K-marginal bias.

Reference graph

Works this paper leans on

16 extracted references · 3 linked inside Pith · cited by 1 Pith paper

[4]

Y. Chen, X. Cheng, J. Niles-Weed, and J. Weare,Convergence of unadjusted langevin in high dimensions: Delo- calization of bias, arXiv preprint arXiv:2408.13115 (2024)

arXiv 2024
[1]

Ahle,Sharp and simple bounds for the raw moments of the binomial and Poisson distributions, Statistics & Probability Letters182(2022), 109306

T.D. Ahle,Sharp and simple bounds for the raw moments of the binomial and Poisson distributions, Statistics & Probability Letters182(2022), 109306

2022
[2]

Altschuler and S

J.M. Altschuler and S. Chewi,Shifted composition II: Shift Harnack inequalities and curvature upper bounds, arXiv preprint arXiv:2401.00071 (2023)

arXiv 2023
[3]

Brunick and S

G. Brunick and S. Shreve,Mimicking an Itˆ o process by a solution of a stochastic differential equation, The Annals of Applied Probability23(2013), no. 4, 1584–1628

2013
[5]

Chewi,Log-concave sampling, Book draft available at https://chewisinho

S. Chewi,Log-concave sampling, Book draft available at https://chewisinho. github. io (2024)

2024
[6]

T. Cui, S. Liu, and X. Tong,Stein ’s method for marginals on large graphical models, arXiv preprint arXiv:2410.11771 (2024)

arXiv 2024
[7]

Durmus and A

A. Durmus and A. Eberle,Asymptotic bias of inexact Markov chain Monte Carlo methods in high dimension, The Annals of Applied Probability34(2024), no. 4, 3435–3468

2024
[8]

Erdos, B

L. Erdos, B. Schlein, and H.-T. Yau,Semicircle law on short scales and delocalization of eigenvectors for Wigner random matrices, The Annals of Probability37(2009), no. 3, 815–852

2009

Show all 16 references

[9]

Lacker,Hierarchies, entropy, and quantitative propagation of chaos for mean field diffusions, Probability and Mathematical Physics4(2023), no

D. Lacker,Hierarchies, entropy, and quantitative propagation of chaos for mean field diffusions, Probability and Mathematical Physics4(2023), no. 2, 377–432

2023
[10]

Lacker and L

D. Lacker and L. Le Flem,Sharp uniform-in-time propagation of chaos, Probability Theory and Related Fields 187(2023), no. 1-2, 443–480

2023
[11]

Lacker, L.C

D. Lacker, L.C. Yeung, and F. Zhou,Quantitative propagation of chaos for non-exchangeable diffusions via first- passage percolation, arXiv preprint arXiv:2409.08882 (2024)

2024 arXiv
[12]

Negrea,Approximations and scaling limits of Markov chains with applications to MCMC and approximate inference, Ph.D

J. Negrea,Approximations and scaling limits of Markov chains with applications to MCMC and approximate inference, Ph.D. thesis, University of Toronto, 2022

2022
[13]

Otto and C

F. Otto and C. Villani,Generalization of an inequality by Talagrand and links with the logarithmic Sobolev inequality, Journal of Functional Analysis173(2000), no. 2, 361–400

2000
[14]

Shou and R

L. Shou and R. van Handel,A localization–delocalization transition for nonhomogeneous random matrices, Journal of Statistical Physics191(2024), no. 2, 26

2024
[15]

Vempala and A

S. Vempala and A. Wibisono,Rapid convergence of the unadjusted Langevin algorithm: Isoperimetry suffices, Advances in neural information processing systems32(2019)

2019
[16]

Villani,Topics in optimal transportation, vol

C. Villani,Topics in optimal transportation, vol. 58, American Mathematical Soc., 2021. Department of Industrial Engineering & Operations Research, Columbia University Email address:daniel.lacker@columbia.edu, fz2329@columbia.edu

2021

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [4]

Y. Chen, X. Cheng, J. Niles-Weed, and J. Weare,Convergence of unadjusted langevin in high dimensions: Delo- calization of bias, arXiv preprint arXiv:2408.13115 (2024)

arXiv 2024

[2] [1]

Ahle,Sharp and simple bounds for the raw moments of the binomial and Poisson distributions, Statistics & Probability Letters182(2022), 109306

T.D. Ahle,Sharp and simple bounds for the raw moments of the binomial and Poisson distributions, Statistics & Probability Letters182(2022), 109306

2022

[3] [2]

Altschuler and S

J.M. Altschuler and S. Chewi,Shifted composition II: Shift Harnack inequalities and curvature upper bounds, arXiv preprint arXiv:2401.00071 (2023)

arXiv 2023

[4] [3]

Brunick and S

G. Brunick and S. Shreve,Mimicking an Itˆ o process by a solution of a stochastic differential equation, The Annals of Applied Probability23(2013), no. 4, 1584–1628

2013

[5] [5]

Chewi,Log-concave sampling, Book draft available at https://chewisinho

S. Chewi,Log-concave sampling, Book draft available at https://chewisinho. github. io (2024)

2024

[6] [6]

T. Cui, S. Liu, and X. Tong,Stein ’s method for marginals on large graphical models, arXiv preprint arXiv:2410.11771 (2024)

arXiv 2024

[7] [7]

Durmus and A

A. Durmus and A. Eberle,Asymptotic bias of inexact Markov chain Monte Carlo methods in high dimension, The Annals of Applied Probability34(2024), no. 4, 3435–3468

2024

[8] [8]

Erdos, B

L. Erdos, B. Schlein, and H.-T. Yau,Semicircle law on short scales and delocalization of eigenvectors for Wigner random matrices, The Annals of Probability37(2009), no. 3, 815–852

2009