arxiv: 2604.22055 · v1 · submitted 2026-04-23 · 🧮 math.NA · cs.NA

Recognition: unknown

A Replica Exchange Markov Chain Monte Carlo Method for Disconnected Implicit Manifolds via Tubular Relaxation

Xuyuan Wang , Donglin Han

Authors on Pith no claims yet

Pith reviewed 2026-05-09 20:30 UTC · model grok-4.3

classification 🧮 math.NA cs.NA

keywords replica exchangeMCMCimplicit manifoldsconstrained samplingtubular relaxationdisconnected componentsMarkov chainsnumerical sampling

0 comments

The pith

A replica exchange MCMC sampler couples constrained and relaxed chains to sample disconnected implicit manifolds.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a Markov chain Monte Carlo method for probability distributions supported on implicit manifolds defined by nonlinear constraints. Standard constrained integrators assume the manifold is connected and therefore cannot move between separate components when that assumption fails. The new approach runs a strict constrained chain on the manifold in parallel with an auxiliary chain that relaxes the constraint inside a tubular neighborhood, allowing the auxiliary chain to cross between components and exchange states with the main chain. The authors prove that the combined sampler obeys detailed balance, is irreducible and ergodic, and converges to the correct target distribution. This extends constrained sampling to a wider set of models arising in molecular and biological dynamics.

Core claim

We propose a replica exchange MCMC framework that couples a constrained chain evolving on the implicit manifold with a relaxed auxiliary chain defined in a tubular neighborhood of the constraint. The relaxed chain enables transitions between disconnected components. We show that the resulting algorithm enables sampling from a broader class of implicit manifolds, including those with disconnected components. We prove that the proposed sampler satisfies detailed balance, irreducibility, ergodicity, and convergence.

What carries the argument

Replica exchange between a constrained chain on the manifold and a relaxed auxiliary chain inside a tubular neighborhood around the constraint, which permits component-crossing transitions while preserving the target measure.

If this is right

The sampler can reach all components of a disconnected manifold, unlike standard constrained Hamiltonian Monte Carlo methods.
The combined chain satisfies detailed balance and therefore has the correct stationary distribution.
The sampler is irreducible and ergodic, guaranteeing convergence to the target measure.
The method applies directly to sampling problems in molecular and biological dynamical systems.
It enlarges the class of implicit manifolds that can be treated by constrained MCMC.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same tubular-relaxation idea could be paired with other constrained integrators beyond RATTLE.
Appropriate tuning of the neighborhood width might allow the method to handle manifolds with more intricate topology than simple disconnections.
Empirical tests on higher-dimensional or more sparsely connected manifolds would map the practical range of relaxation parameters that still preserve convergence.
The framework might combine with Hamiltonian dynamics in the auxiliary chain to improve mixing rates in high-dimensional settings.

Load-bearing premise

The tubular neighborhood must be wide enough and the relaxation gentle enough for the auxiliary chain to reach and cross between disconnected components without destroying convergence.

What would settle it

Construct a simple disconnected manifold such as two separate circles defined by a nonlinear constraint, run the sampler, and check whether it visits both components with frequencies matching the target distribution.

Figures

Figures reproduced from arXiv: 2604.22055 by Donglin Han, Xuyuan Wang.

**Figure 1.** Figure 1: Comparison of Algorithm 2 and Algorithm 4 on a synthetic example with four disconnected elliptical components. Having demonstrated that Algorithm 4 enables efficient exploration across disconnected components, we now examine the bias introduced by replacing the exact Jacobian determinant in the Metropolis–Hastings acceptance probability with its Gram approximation in equation (14), as analyzed in Theorem… view at source ↗

**Figure 2.** Figure 2: Numerical validation of Theorem 4.5 using exact and Gram-approximated Jacobian determinants in Algorithm 4. 5.2 Infectious Disease Models The next numerical example arises from Bayesian inverse problems in infectious disease modeling [Roda, 2020, Wang, 2026]. In this setting, one begins with a specified and evaluable forward model, typically derived from a system of differential equations describing the dy… view at source ↗

**Figure 3.** Figure 3: SIR model with two infectious compartments. Left: governing differential equations. [PITH_FULL_IMAGE:figures/full_fig_p036_3.png] view at source ↗

**Figure 4.** Figure 4: Sampling results on the disconnected implicit manifold [PITH_FULL_IMAGE:figures/full_fig_p038_4.png] view at source ↗

read the original abstract

Markov chain Monte Carlo (MCMC) methods provide powerful framework for sampling unknown probability measures across a wide range of scientific applications. In some settings, the target distribution is supported on a lower-dimensional submanifold of Euclidean space defined by nonlinear constraints, motivating the development of constrained Hamiltonian Monte Carlo (CHMC) methods. Most existing CHMC algorithms rely on the assumption that the implicit manifold is connected, allowing local constrained integrators such as RATTLE to explore the posterior ergodically. In practice, this assumption is occasionally violated due to complex geometric structures induced by nonlinear constraints of a model. We propose a replica exchange MCMC framework that couples a constrained chain evolving on the implicit manifold with a relaxed auxiliary chain defined in a tubular neighborhood of the constraint. The relaxed chain enables transitions between disconnected components. We show that the resulting algorithm enables sampling from a broader class of implicit manifolds, including those with disconnected components. We prove that the proposed sampler satisfies detailed balance, irreducibility, ergodicity, and convergence. We also demonstrate its effectiveness on examples from molecular and biological dynamical systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives a replica-exchange MCMC construction for sampling disconnected implicit manifolds by relaxing into a tubular neighborhood, with asserted proofs of the usual properties, though the conditions needed for the auxiliary chain to actually connect components are not spelled out in the abstract.

read the letter

This paper's main contribution is a replica exchange framework for MCMC on implicit manifolds that may be disconnected. It runs a standard constrained Hamiltonian Monte Carlo chain on the manifold itself alongside an auxiliary chain that operates in a tubular neighborhood where the constraint is relaxed. The exchange allows the auxiliary chain to move between separate components and bring the constrained chain along. What stands out is how they target a practical problem that arises in molecular dynamics and biological models when nonlinear constraints split the space into disconnected pieces. Standard CHMC assumes connectivity for ergodicity, so this extension makes sense. They also provide numerical examples to illustrate the method working on those kinds of systems. The proofs for detailed balance, irreducibility, ergodicity, and convergence are asserted directly. That is better than many papers that skip theory entirely. The stress-test concern is worth noting: irreducibility depends on the auxiliary chain actually being able to cross between components, which requires the tubular neighborhood to be sufficiently wide or the relaxation to be appropriate for the separation between components. Without stated bounds or geometric conditions on the constraint function and the tube radius, it is hard to know how general the result is. If the full paper derives explicit requirements or shows robustness, that would address it; from the abstract it seems those might be implicit. The citation pattern looks standard for the area, referencing CHMC methods without obvious loops. Overall, this is for people already working on manifold-constrained sampling or replica exchange methods. It builds incrementally on existing CHMC literature rather than overhauling the field. I would send this to peer review. The problem is well-motivated, the method is clearly described at a high level, and the theoretical claims are checkable by referees familiar with MCMC on manifolds.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a replica exchange MCMC sampler for probability measures supported on implicit manifolds defined by nonlinear constraints. It couples a constrained chain (e.g., via RATTLE) that evolves exactly on the manifold with an auxiliary chain whose potential is relaxed inside a tubular neighborhood of the constraint surface; the auxiliary chain is intended to enable transitions between disconnected components. The authors prove that the resulting Markov chain satisfies detailed balance, is irreducible and ergodic, and converges to the target distribution, and they illustrate the method on molecular and biological examples.

Significance. If the proofs are rigorous, the work meaningfully extends constrained Hamiltonian Monte Carlo to the practically relevant case of disconnected manifolds. The tubular-relaxation construction supplies a concrete mechanism for inter-component moves while retaining exact constrained dynamics on the manifold itself. Explicit proofs of detailed balance, irreducibility, ergodicity, and convergence constitute a clear theoretical strength.

major comments (2)

[§4.2, Theorem 4.1] §4.2, Theorem 4.1 (irreducibility): the argument that the auxiliary chain can move between disconnected components rests on the tubular neighborhood being wide enough for the relaxed dynamics to traverse the gaps. No quantitative lower bound on tube radius (or on the relaxation strength) relative to the separation of the components or the geometry of the constraint map is stated. Without such a condition the transition kernel may remain reducible on some manifolds, undermining the central claim that the sampler works for arbitrary disconnected implicit manifolds.
[§3.3] §3.3 (convergence): the proof of ergodicity and convergence invokes the irreducibility result of §4.2 together with the standard Harris-recurrence argument for replica-exchange kernels. Because the irreducibility step lacks an explicit geometric hypothesis, the convergence statement inherits the same gap; a counter-example manifold on which the tube fails to connect components would falsify the claim.

minor comments (2)

[Preliminaries] The notation for the tube radius and the relaxation potential is introduced only in the algorithmic section; moving the definitions to the preliminaries would improve readability.
[Figure 1] Figure 1 (schematic of the tubular neighborhood) would benefit from an explicit label indicating the distance between the two disconnected components relative to the tube radius.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments, which help strengthen the theoretical foundations of the work. We address each major comment below.

read point-by-point responses

Referee: [§4.2, Theorem 4.1] §4.2, Theorem 4.1 (irreducibility): the argument that the auxiliary chain can move between disconnected components rests on the tubular neighborhood being wide enough for the relaxed dynamics to traverse the gaps. No quantitative lower bound on tube radius (or on the relaxation strength) relative to the separation of the components or the geometry of the constraint map is stated. Without such a condition the transition kernel may remain reducible on some manifolds, undermining the central claim that the sampler works for arbitrary disconnected implicit manifolds.

Authors: We agree that an explicit quantitative hypothesis is required to guarantee irreducibility on arbitrary disconnected manifolds. The current proof of Theorem 4.1 implicitly assumes the tubular neighborhood is wide enough for the auxiliary chain to traverse gaps between components, but does not state a precise lower bound in terms of inter-component distance or the geometry of the constraint map. We will revise the theorem statement to include such a bound (derived from the minimal separation of components and the Lipschitz constant of the defining constraint function) and update the proof to verify that the relaxed dynamics connect components under this condition. This makes the irreducibility claim rigorous while preserving the algorithm and its applicability to the manifolds arising in the molecular and biological examples. revision: yes
Referee: [§3.3] §3.3 (convergence): the proof of ergodicity and convergence invokes the irreducibility result of §4.2 together with the standard Harris-recurrence argument for replica-exchange kernels. Because the irreducibility step lacks an explicit geometric hypothesis, the convergence statement inherits the same gap; a counter-example manifold on which the tube fails to connect components would falsify the claim.

Authors: We concur that the ergodicity and convergence results in §3.3 rest on the irreducibility established in Theorem 4.1. Once the explicit geometric condition on tube radius is incorporated into Theorem 4.1, the Harris-recurrence argument will apply directly under the same hypothesis. We will update the statements and proofs in §3.3 to reference this condition explicitly, thereby closing the gap and ensuring the convergence claim holds precisely for the class of manifolds satisfying the tube-radius requirement. revision: yes

Circularity Check

0 steps flagged

No circularity: proofs derive from algorithmic construction and standard MCMC theory

full rationale

The paper introduces a replica-exchange MCMC coupling a constrained chain on the implicit manifold with a relaxed auxiliary chain in a tubular neighborhood. It claims to prove detailed balance, irreducibility, ergodicity, and convergence for the resulting sampler. These properties follow from the explicit transition kernels defined by the replica-exchange mechanism and the relaxation inside the tube, together with standard arguments from Markov chain theory (e.g., Harris recurrence or Foster-Lyapunov drift conditions). No equation reduces a claimed prediction or uniqueness result to a fitted parameter or to a prior self-citation by construction. The abstract and description contain no self-definitional steps, no renaming of known empirical patterns, and no load-bearing reliance on unverified self-citations. The derivation is therefore self-contained against external benchmarks of MCMC convergence.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are stated. The tubular neighborhood and relaxation schedule are algorithmic choices whose precise definitions and any hidden tuning parameters are not visible.

pith-pipeline@v0.9.0 · 5489 in / 1197 out tokens · 17848 ms · 2026-05-09T20:30:46.880576+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

39 extracted references · 4 canonical work pages · 1 internal anchor

[1]

The Journal of chemical physics , volume=

Bayesian uncertainty quantification and propagation in molecular dynamics simulations: a high performance computing framework , author=. The Journal of chemical physics , volume=. 2012 , publisher=

2012
[2]

Reviews of Modern Physics , volume=

Bayesian inference in physics , author=. Reviews of Modern Physics , volume=. 2011 , publisher=

2011
[3]

Infectious Disease Modelling , volume=

Bayesian inference for dynamical systems , author=. Infectious Disease Modelling , volume=. 2020 , publisher=

2020
[4]

ZAMM-Journal of Applied Mathematics and Mechanics/Zeitschrift f

A constrained hybrid Monte-Carlo algorithm and the problem of calculating the free energy in several variables , author=. ZAMM-Journal of Applied Mathematics and Mechanics/Zeitschrift f. 2005 , publisher=

2005
[5]

Journal of Computational Physics , volume=

Monte Carlo on manifolds in high dimensions , author=. Journal of Computational Physics , volume=. 2024 , publisher=

2024
[6]

PLoS computational biology , volume=

Structural identifiability of dynamic systems biology models , author=. PLoS computational biology , volume=. 2016 , publisher=

2016
[7]

arXiv preprint arXiv:2511.12847 , year=

Identification-aware Markov chain Monte Carlo , author=. arXiv preprint arXiv:2511.12847 , year=

work page arXiv
[8]

Bulletin of Mathematical Biology , volume=

Bayesian Identifiability Analysis for Infectious Disease Models: Parameter Reduction and Model Selection , author=. Bulletin of Mathematical Biology , volume=. 2026 , publisher=

2026
[9]

Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

Manifold lifting: scaling Markov chain Monte Carlo to the vanishing noise regime , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=
[10]

SIAM/ASA Journal on Uncertainty Quantification , volume=

Sampling parameters of ordinary differential equations with constrained Langevin dynamics , author=. SIAM/ASA Journal on Uncertainty Quantification , volume=. 2025 , publisher=

2025
[11]

2010 , publisher=

Free energy computations: A mathematical perspective , author=. 2010 , publisher=

2010
[12]

Artificial intelligence and statistics , pages=

A family of MCMC methods on implicitly defined manifolds , author=. Artificial intelligence and statistics , pages=. 2012 , organization=

2012
[13]

Numerische Mathematik , volume=

Hybrid Monte Carlo methods for sampling probability measures on submanifolds , author=. Numerische Mathematik , volume=. 2019 , publisher=

2019
[14]

Communications on Pure and Applied Mathematics , volume=

Monte Carlo on manifolds: sampling densities and integrating functions , author=. Communications on Pure and Applied Mathematics , volume=
[15]

arXiv preprint arXiv:2206.11267 , year=

Neural implicit manifold learning for topology-aware density estimation , author=. arXiv preprint arXiv:2206.11267 , year=

work page arXiv
[16]

velocity

Rattle: A “velocity” version of the shake algorithm for molecular dynamics calculations , author=. Journal of computational Physics , volume=. 1983 , publisher=

1983
[17]

Advances in Neural Information Processing Systems , volume=

Local-global mcmc kernels: the best of both worlds , author=. Advances in Neural Information Processing Systems , volume=
[18]

Journal of Computational Chemistry , volume=

Collision-free poisson motion planning in ultra high-dimensional molecular conformation spaces , author=. Journal of Computational Chemistry , volume=. 2018 , publisher=

2018
[19]

Current opinion in systems biology , volume=

On structural and practical identifiability , author=. Current opinion in systems biology , volume=. 2021 , publisher=

2021
[20]

Chemical physics letters , volume=

Replica-exchange molecular dynamics method for protein folding , author=. Chemical physics letters , volume=. 1999 , publisher=

1999
[21]

SIAM Journal on Mathematics of Data Science , volume =

Breiding, Paul and Marigliano, Orlando , title =. SIAM Journal on Mathematics of Data Science , volume =
[22]

Handbook of markov chain monte carlo , pages=

MCMC using Hamiltonian dynamics , author=. Handbook of markov chain monte carlo , pages=. 2011 , publisher=

2011
[23]

The american statistician , volume=

Understanding the metropolis-hastings algorithm , author=. The american statistician , volume=. 1995 , publisher=

1995
[24]

2006 , publisher=

Algorithms in real algebraic geometry , author=. 2006 , publisher=

2006
[25]

2023 , publisher=

An introduction to smooth manifolds , author=. 2023 , publisher=

2023
[26]

Journal of computational chemistry , volume=

Algorithms for constrained molecular dynamics , author=. Journal of computational chemistry , volume=. 1995 , publisher=

1995
[27]

2005 , publisher=

Probability measures on metric spaces , author=. 2005 , publisher=

2005
[28]

2009 , publisher=

Optimal transport: old and new , author=. 2009 , publisher=

2009
[29]

arXiv preprint arXiv:1704.03338 , year=

Continuously tempered hamiltonian monte carlo , author=. arXiv preprint arXiv:1704.03338 , year=

work page arXiv
[30]

International Conference on Machine Learning , pages=

Non-convex learning via replica exchange stochastic gradient mcmc , author=. International Conference on Machine Learning , pages=. 2020 , organization=

2020
[31]

Application to a chikungunya model , author=

A distribution input--output polynomial approach for estimating parameters in nonlinear models. Application to a chikungunya model , author=. Journal of Computational and Applied Mathematics , volume=. 2018 , publisher=

2018
[32]

A Tutorial on Structural Identifiability of Epidemic Models Using StructuralIdentifiability.jl

A tutorial on structural identifiability of epidemic models using structuralidentifiability. jl , author=. arXiv preprint arXiv:2505.10517 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[33]

Communications in Nonlinear Science and Numerical Simulation , volume=

A two-strain avian--human influenza model with environmental transmission: stability analysis and optimal control strategies , author=. Communications in Nonlinear Science and Numerical Simulation , volume=. 2024 , publisher=

2024
[34]

2023 , publisher=

Understanding molecular simulation: from algorithms to applications , author=. 2023 , publisher=

2023
[35]

2023 , publisher=

Statistical mechanics: theory and molecular simulation , author=. 2023 , publisher=

2023
[36]

Bernoulli , volume=

Perturbation theory for Markov chains via Wasserstein distance , author=. Bernoulli , volume=. 2018 , publisher=

2018
[37]

Transactions of the American Mathematical Society , volume=

Curvature measures , author=. Transactions of the American Mathematical Society , volume=. 1959 , publisher=

1959
[38]

2018 , publisher=

Introduction to Riemannian manifolds , author=. 2018 , publisher=

2018
[39]

2013 , publisher=

Convergence of probability measures , author=. 2013 , publisher=

2013