REVIEW 2 major objections 6 minor 1 cited by
Foundations of locally-balanced Markov processes
T0 review · 2 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash
Pith's one-line read The paper establishes that locally-balanced Markov jump processes are well-posed, reversible and ergodic on general state spaces, and can be uniformly ergodic where overdamped Langevin diffusion is not.
desk verdict Solid, mostly rigorous foundations paper for locally-balanced Markov jump processes; central theorems hold and the localized proof errors are fixable. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the balancing function g: R≥0 → R≥0 with g(1)=1 and g(t)=t g(1/t), combined with the Radon–Nikodym derivative t(x,y) of the time-reversed proposal against the forward one. The balancing identity makes the jump kernel J(x,dy)=g(t(x,y))γ(x,dy) reversible with respect to π; the monotonicity assumption on g supplies the two-sided bounds min(1,t)≤g(t)≤max(1,t), which yield finite jump rates, non-explosivity, integrability of the rate, and the spectral-gap comparisons; and for the diffusion limit, the identity g′(1)=1/2 (a consequence of the balancing identity and g(1)=1) is what makes the limiting drift exactly (1/2)∇log π.
What would settle it
Exhibit a balancing function g that satisfies g(1)=1 and g(t)=t g(1/t) but is not monotone, together with a target π and proposal γ for which λ(x)=∫g(t(x,y))γ(x,dy)=∞ on a set of positive π-measure; this would show the monotonicity assumption is genuinely load-bearing and break the well-posedness theorem as stated. Alternatively, for the uniform-ergodicity claim, simulate the embedded nearest-neighbour chain of Theorem 3.1 for π(n)∝exp(−n^a), a in (1,2), with g(t)=t; if the expected hitting time of a compact set from very large n does not stay bounded as n→∞, the uniform-ergodicity conclusion would fail.
Extended reading notes
Core claim
The central claim is that the locally-balanced Markov jump process, defined through a balancing function g, a base kernel γ, and target π, is a genuine Markov jump process: under Assumptions 2.1–2.4 it is non-explosive, π-reversible, and ergodic from π-almost every starting state, and from every starting state in the support of π under Assumption 2.4. The key structural fact is that the jump kernel J(x, dy) = g(t(x, y))γ(x, dy), with t(x, y) the Radon–Nikodym derivative of the time-reversed proposal against the forward one, satisfies detailed balance because g(t) = t g(1/t), so π(dx)J(x, dy) is symmetric. With g non-decreasing and continuous, the paper proves min(1,t) ≤ g(t) ≤ max(1,t), which makes the jump rate λ integrable under π and the process non-explosive. It further claims: a spectral-gap equivalence between LBMJPs with bounded g and Metropolis–Hastings algorithms with the same proposal; comparison theorems for unbounded g; uniform ergodicity on unbounded spaces for targets like π(n) ∝ exp(−n^a) with a in (1,2) when g grows, a regime where overdamped Langevin diffusion is not uniformly ergodic; and a weak diffusion limit to the overdamped Langevin diffusion as the proposal variance goes to zero with time rescaled by $σ_n^{{−2}}$.
Load-bearing premise
The load-bearing premise is Assumption 2.2: the balancing function is non-decreasing and continuous; without monotonicity, the paper's two-sided bounds on g fail, and the jump rate can in principle be infinite, which would make the process itself undefined.
Editorial extensions
If this is right
- If a Metropolis–Hastings chain with proposal γ and target π has a positive spectral gap, then the corresponding LBMJP with any non-decreasing balancing function and the same γ also has a positive spectral gap, so LBMJPs inherit a large body of geometric-ergodicity results.
- With unbounded g, LBMJPs can be uniformly ergodic on unbounded state spaces for targets with tails that are not sub-Gaussian, such as π(n) ∝ exp(−n^a) with a in (1,2), a regime where the overdamped Langevin diffusion is not uniformly ergodic.
- In the small-jump limit with time rescaled, a LBMJP converges weakly to the overdamped Langevin diffusion, so the behaviour of a finely tuned LBMJP approaches that of a Langevin sampler.
- The embedded discrete-time chain of a LBMJP can be used for Monte Carlo estimation either through continuous-time averages or through importance-tempered averages, with the latter having lower asymptotic variance.
- The same balancing construction extends to non-reversible processes through skewed or modified detailed balance, giving a route to non-reversible locally-balanced samplers.
Reading between the lines
- Combining the uniform-ergodicity result on a fixed grid with the diffusion limit on the same grid suggests a genuine order-of-limits phenomenon: a fixed fine grid can yield uniform ergodicity for a heavy-tailed target, while the continuum limit loses it, so the choice of grid spacing relative to the time horizon is a substantive modelling decision rather than a numerical artefact.
- The monotonicity of g is used heavily, but the reversibility property itself follows from the balance identity alone; a plausible extension is that some non-monotone balancing functions also yield well-posed processes, and the paper explicitly leaves this case open in Remark 2.4.
- The spectral-gap comparison with Metropolis–Hastings suggests a practical recipe: take any geometrically ergodic Metropolis–Hastings sampler, replace its accept-reject step by the continuous-time locally-balanced jump kernel with the same proposal, and retain exponential ergodicity without extra dimension-dependent tuning.
- The variance reduction of importance tempering over continuous-time averaging, proved in the paper's setting, likely also applies to the non-reversible extensions sketched in Section 5.2, since the embedded-chain structure is preserved there, although the paper does not prove this.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper develops a mathematical foundation for locally-balanced Markov jump processes (LBMJPs) on Polish state spaces. For a target probability π and a Markov proposal kernel γ, the process has jump rate λ(x)=∫ g(t(x,y))γ(x,dy) and jump distribution proportional to g(t(x,y))γ(x,dy), where t(x,y) is the Radon-Nikodym ratio π(dy)γ(y,dx)/(π(dx)γ(x,dy)) and g is a balancing function. Under continuity and monotonicity of g (Assumption 2.2) plus irreducibility and support conditions, the paper proves λ finite, non-explosivity, π-reversibility, ergodic theorems, weak Feller property, and a weak generator. It then shows spectral-gap comparison with Metropolis-Hastings for bounded g, comparison theorems for unbounded g, uniform ergodicity on N for targets with exponential-type ratio tails such as π(n)∝exp(-n^a), a∈(1,2), and a diffusion limit to overdamped Langevin when Gaussian proposal variance σ_n^2→0 with time sped up by σ_n^{-2}. The proofs use standard tools from Markov chain theory and piecewise deterministic Markov processes, with several technical arguments placed in appendices.
Significance. If the results stand, this paper supplies a rigorous common framework for a class of continuous-time samplers that is increasingly used in Monte Carlo for discrete or non-smooth targets. The spectral-gap comparison with Metropolis-Hastings is valuable because it transfers a large body of known results; the uniform ergodicity example for non-sub-Gaussian targets is a genuinely interesting phenomenon that contrasts with Langevin diffusions; and the diffusion limit connects LBMJPs to optimal-scaling literature. The presentation is careful and the reliance on external results (Tierney, Norris, Meyn-Tweedie, Down et al., Monmarché et al.) is made explicit. The restriction to non-decreasing balancing functions is a stated limitation rather than a hidden assumption, and the authors are candid about it. I found no error that changes the main theorems; the issues below are localized and repairable.
major comments (2)
- [Section 2.3, proof of Proposition 2.3] The construction of the process in the weak-Feller proof sets τ^x_1 = λ(x)E_1 with E_1 ~ Exp(1). Since E_1/λ(x) is Exp(λ(x))-distributed and λ(x)E_1 is Exp(1/λ(x))-distributed, the construction inverts the required waiting-time scaling. Replacing λ(x) by 1/λ(x) in the displayed formulas repairs the proof without changing the conclusion, because continuity of λ and λ>0 suffice for the argument.
- [Section 3.2, proof of Proposition 3.5] The statement that the Down et al. drift condition is 'trivially satisfied' by setting V_T≡1, β(s)=0 and b=1 is not the standard form of the condition, which involves an indicator of a petite set and typically requires β(T)<1 with a non-trivial V. The argument can be made rigorous by using the uniform minorization in (24) to show that the whole space is petite, or by constructing a Lyapunov function from the uniform hitting-time bound, but the proof should spell this out.
minor comments (6)
- [Appendix C] Appendix C contains an incorrect numerical value: direct differentiation of b(x)=log g(e^x) gives b''(0)=g'(1)+g''(1)-g'(1)^2=1/4+g''(1), not 1/2+g''(1)-g''(1)^2 as displayed; since only boundedness of b'' is used in the proof of Theorem 4.1, the conclusion is unaffected, but the formula should be corrected.
- [Section 2.3] The definitions of strong and weak Feller are conflated; as written, the sentence 'We will refer to a Markov process for which ... as strong Feller if ...' does not match the standard definitions, and the weak Feller property should be stated as P_t f∈C_b(E) for every f∈C_b(E).
- [Remark 2.6] The final sentence of Remark 2.6 invokes 'Assumption 2.1' where the surrounding argument concerns Assumption 2.4; this should be corrected.
- [Section 3.2.1, Example 1] The kernel γ(x,·)=1/2(δ_{x-1}+δ_{x+1}) is not defined at x=0 on the state space N; a boundary convention should be stated.
- [Appendix B, proof of Proposition B.1] The coupling of Y and \tilde Y by 'using the same exp(1) random variables' is not explicit; since the jump rates differ, the construction should be described (e.g., via Poisson thinning or a common Poisson clock with acceptance probabilities) to justify the stochastic domination used to bound E_{n+1}[h_n].
- [Section 2.3, proof of Proposition 2.3] The phrase 'since T^y_n∈ N for all y' appears to be a typo: the intended meaning is that T^y_n is continuous in y, so that T^y_n=T^x_n for y sufficiently close to x.
Circularity Check
No significant circularity: central results are proved from stated assumptions and external theorems; self-citations are contextual and non-load-bearing.
full rationale
The derivation chain is self-contained relative to its assumptions. Definition 2.2 constructs the LBMJP from a balancing function g, a base kernel gamma, and a target pi; the reversibility and invariance results (Proposition 2.1 and Theorem 2.1(b)) follow from the balancing identity g(t)=t g(1/t), not from an imposed demand that pi be stationary. Non-explosivity, ergodicity, the Feller property, the weak generator, spectral-gap comparisons, uniform ergodicity, and the diffusion limit are all proved from Assumptions 2.1-2.5 using external standard results (Tierney 1998, Norris 1997, Asmussen and Glynn 2011, Davis 1984, Ethier and Kurtz 2009, Meyn and Tweedie, Down et al. 1995, Monmarché et al. 2022). The Metropolis-Hastings comparison in Proposition 3.2 is a genuine two-sided spectral-gap bound, not a restatement of one quantity as the other. In Theorem 4.1 the Langevin drift arises from the Taylor expansion of the balancing function and the identity g'(1)=1/2, which follows from g(t)=t g(1/t); no drift coefficient is fitted. Self-citations, such as Zanella 2020, Livingstone and Zanella 2022, and Vogrinc et al. 2022, appear in historical remarks, examples, and discussion sections, and none is load-bearing for the main theorems. Stated limitations, such as Remark 2.4 leaving non-monotone g to future work, are explicit assumptions and not hidden input-output identifications. Localized proof defects, including the tau scaling typo in Proposition 2.3 and the b''(0) value in Appendix C, are technical issues rather than circular reasoning. Therefore the paper exhibits no significant circularity.
Assumptions & free parameters
assumptions (9)
- standard math E is a Polish space and the Radon-Nikodym derivative t(x,y) is well-defined via Tierney's Proposition 1.
- domain assumption Assumption 2.1: lambda(x) > 0 for all x in supp(pi).
- domain assumption Assumption 2.2: g is non-decreasing and continuous.
- domain assumption Assumption 2.3: Gamma_g is phi-irreducible.
- domain assumption Assumption 2.4: Gamma_g(x,.) is absolutely continuous with respect to pi for all x in supp(pi).
- domain assumption Assumption 2.5: gamma has a continuous representation and t is continuous.
- domain assumption Assumption 3.1: g(t) >= t^a for t >= 1 and pi(n)/pi(n+1) >= exp(a beta n^{beta-1}) for large n.
- domain assumption Assumption 4.1: pi in C^3, pi bounded away from 0, g in C^2, and the Hessian of log pi is uniformly bounded.
- standard math External theorems used as black boxes: Norris 1997 Theorem 2.3.2, Asmussen-Glynn 2011 Theorem 1, Davis 1984 Theorem 5.5, Feller 1991, Serfozo 2005, Meyn-Tweedie 1992/2009, Down et al. 1995, Monmarche et al. 2022 Theorem 8.1.
Cite this review
Pith. "Pith review of Foundations of locally-balanced Markov processes." pith.science (2026). https://pith.science/paper/FN6RIUE3
@misc{pith2026250413322,
author = {Pith},
title = {Pith review of: Foundations of locally-balanced Markov processes},
year = {2026},
howpublished = {\url{https://pith.science/paper/FN6RIUE3}},
note = {Machine review of arXiv:2504.13322}
}
read the original abstract
We formally introduce and study locally-balanced Markov jump processes (LBMJPs) defined on a general state space. These continuous-time stochastic processes with a user-specified limiting distribution are designed for sampling in settings involving discrete parameters and/or non-smooth distributions, addressing limitations of other processes such as the overdamped Langevin diffusion. The paper establishes the well-posedness, non-explosivity, and ergodicity of LBMJPs under mild conditions. We further explore regularity properties such as the Feller property and characterise the weak generator of the process. We then derive conditions for exponential ergodicity via spectral gaps and establish comparison theorems for different balancing functions. In particular we show an equivalence between the spectral gaps of Metropolis--Hastings algorithms and LBMJPs with bounded balancing function, but show that LBMJPs can exhibit uniform ergodicity on unbounded state spaces when the balancing function is unbounded, even when the limiting distribution is not sub-Gaussian. We also establish a diffusion limit for an LBMJP in the small jump limit, and discuss applications to Monte Carlo sampling and non-reversible extensions of the processes.
Forward citations
Cited by 1 Pith paper
-
From Minimax Optimal Importance Sampling to Uniformly Ergodic Importance-tempered MCMC
A minimax analysis identifies the optimal importance-sampling proposal for atomic targets, and an exact uniform ergodicity criterion is proved for importance-tempered random-walk Metropolis on polynomial-tail targets.
Reference graph
Works this paper leans on
-
[1]
Peskun–Tierne y ordering for Markovian Monte Carlo: beyond the reversible scenario
Christophe Andrieu and Samuel Livingstone. Peskun–Tierne y ordering for Markovian Monte Carlo: beyond the reversible scenario. The Annals of Statistics , 49(4):1958–1981,
work page 1958
-
[3]
Let Ln and L be the weak generators of Sn and S respectively. Recall from Theorem 2.3 that for f∈ C ∞ c (Rd), Lnf (x) = σ−2 n EY ∼N (x,σ 2nId) [(f (Y )− f (x)) g◦ t(x, Y )] = σ−1 n EZ∼N (0,I d) [(f (x + σnZ)− f (x)) exp{b (φ(x + σnZ)− φ(x))}] , while it is well known (see e.g. Gelman et al. [1997]) that Lf (x) = 1 2∇φ(x)∇f (x) + 1 2 ∆ f (x). 28 SAMUEL LIV...
work page 1997
-
[7]
Luke Hardcastle, Samuel Livingstone, and Gianluca Baio. Av eraging polyhazard models using Piecewise deterministic Monte Carlo with applications to d ata with long-term survivors. arXiv preprint arXiv:2406.14182,
-
[11]
doi: https://doi.org/10.1007/s40072-022-00247-9 . J. R. Norris. Markov Chains . Cambridge Series in Statistical and Probabilistic Mathem atics. Cambridge University Press,
-
[12]
doi: 10.1017/CBO978051 1810633. A. Pazy. Semigroups of linear operators and applications to partial differential equations, volume
-
[15]
ISSN 1464-3510. doi: 10.1093/biomet/asac056. Tatiana Xifara, Chris Sherlock, Samuel Livingstone, Simon Byrne, and Mark Girolami. Langevin diffusions and the Metropolis-adjusted Langevin algorithm. Statistics & Probability Letters , 91: 14–19,
-
[16]
(36) We then have the following
Assume further that nb(n) n→∞−−−→0. (36) We then have the following. Proposition B.1. Assume that Assumption B.1 holds. Then for any compact set C⊂ N, lim sup N →∞ EN [hC ] <∞ ⇐⇒ ∞∑ n=1 a(n) <∞. (37) Proof of Proposition B.1. Let C⊂ N be a compact set and let k = sup C. We first observe that for any N > k , for the LBMJP with γ as in (25), hC = hk. We will...
work page 2006
-
[1992]
doi: 10.2307/1427479. S. P. Meyn and R. L. Tweedie. Stability of Markovian processe s II: continuous-time processes and sampled chains. Advances in Applied Probability , 25(3):487–517,
Show all 17 references
-
[1993]
Sean Meyn and Richard L
doi: 10.2307/1427521. Sean Meyn and Richard L. Tweedie. Markov Chains and Stochastic Stability . Cambridge Mathe- matical Library. Cambridge University Press, 2 edition, 20
-
[1997]
Will Grathwohl, Kevin Swersky, Milad Hashemi, David Duvena ud, and Chris Maddison
doi: 10.1214/aoap/1034625254. Will Grathwohl, Kevin Swersky, Milad Hashemi, David Duvena ud, and Chris Maddison. Oops i took a gradient: Scalable sampling for discrete distributi ons. In International Conference on Machine Learning, pages 3831–3841. PMLR,
-
[2000]
doi: 10.1214/aoap/ 1019487508
ISSN 1050-5164. doi: 10.1214/aoap/ 1019487508. Peter J Diggle. Modeling infectious disease distributions : Applications of point process methods. In Handbook of infectious disease data analysis , pages 387–409. Chapman and Hall/CRC,
-
[2012]
Accelerated sampl ing on discrete spaces with non- reversible Markov processes
Samuel Power and Jacob Vorstrup Goldman. Accelerated sampl ing on discrete spaces with non- reversible Markov processes. arXiv preprint arXiv:1912.04681 ,
1912 arXiv
-
[2017]
Bierkens, G
J. Bierkens, G. O. Roberts, and Pierre-Andr´ e Zitt. Ergodic ity of the zigzag process. The Annals of Applied Probability , 29(4):2266 – 2301, 2019a. doi: 10.1214/18-AAP1453. Joris Bierkens, Paul Fearnhead, and Gareth Roberts. The Zig -Zag process and super-efficient sampling fo...
-
[2018]
Dimension-free relaxation t imes of informed mcmc samplers on discrete spaces
Hyunwoong Chang and Quan Zhou. Dimension-free relaxation t imes of informed mcmc samplers on discrete spaces. arXiv preprint arXiv:2404.03867 ,
-
[2021]
doi: 10.1214/20-AIHP1125. S. N. Ethier and T. G. Kurtz. Markov processes: characterization and convergence . John Wiley & Sons,
-
[2022]
doi: 10.1111/rss b.12482
ISSN 1369-7412. doi: 10.1111/rss b.12482. I. Lytras and P. Mertikopoulos. Tamed Langevin sampling und er weaker conditions,
-
[2023]
Nonreversible MCMC from conditional invertible tra nsforms: a complete recipe with convergence guarantees
Achille Thin, Nikita Kotelevskii, Christophe Andrieu, Ala in Durmus, Eric Moulines, and Maxim Panov. Nonreversible MCMC from conditional invertible tra nsforms: a complete recipe with convergence guarantees. arXiv preprint arXiv:2012.15550 ,
2012 arXiv
Reviewed August 16, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.