The nonequilibrium statistical mechanics of Markov interacting particles

Dalton A R Sakthivadivel

REVIEW 3 major objections 6 minor 27 references

The nonequilibrium statistical mechanics of Markov interacting particles

T0 review · 3 major / 6 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read A boundary path screens off the exterior from the interior exactly when the boundary's conditional likelihood separates into a product of an exterior factor and an interior factor—and this single test unifies likelihood inference, control e

desk verdict Useful synthesis of path-space conditional independence; the new thermodynamic bounds have a concrete Girsanov metric error and the domination assumption is understated. read the letter →

arxiv 2607.13391 v1 pith:2JEDQTLC submitted 2026-07-15 cond-mat.stat-mech math-phmath.MPmath.PRnlin.AO

Dalton A R Sakthivadivel This is my paper · ORCID

classification cond-mat.stat-mechmath-phmath.MPmath.PRnlin.AO MSC 60H0760J6060J2560G1560G2262F1582C31

keywords pathspaceMarkovboundaryconditionalindependenceRadon–NikodymseparabilityGirsanovlikelihoodFöllmerentropyidentitymutualinformationnonequilibriumsteadystatesensor-actuatorsystem

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 asks when, in a stochastic system split into exterior, boundary, and interior variables, observing the entire boundary history makes the exterior and interior histories independent of each other. It proves that this path-space Markov boundary property is equivalent, under a domination assumption, to the multiplicative separation of a Radon–Nikodym density of the conditioned law with respect to a clamped product reference law. The same condition reappears as additive separation of a path log-likelihood, as a zero in the cross precision block for Gaussian path laws, and as a split of the conditional generator for Markov diffusions. When the condition fails, the deviation is measured by conditional mutual information, which the paper connects to the excess free energy or work needed to enforce a factorised, separated architecture. The value is a common probabilistic language for deciding when a boundary genuinely decouples a system from its environment.

What carries the argument

The central object is a regular conditional probability on a Polish path space, giving rigorous meaning to conditioning one trajectory on another. The load-bearing identity is the Radon–Nikodym separability criterion: a probability measure is a product measure iff its density with respect to a product reference factorizes multiplicatively. In the Itô setting, Girsanov's theorem computes the log-likelihood and its expectation is the quadratic control energy appearing in the entropy identity for path measures. The factorisation gap is the conditional mutual information, defined as a fibrewise relative entropy, which also serves as the thermodynamic excess free energy of a factorised architectu

What would settle it

Construct a system in which the boundary path is a deterministic function of the exterior and interior histories, such as B_t = Y_t + X_t, with the driving noises chosen so that conditioning on B leaves both Y and X fully determined. Then Law(Y[0,T], X[0,T] | B[0,T]=b) is a Dirac measure concentrated on a single curve, hence singular with respect to any non-atomic product reference. In this case Theorem 5.3 does not apply, and one can directly check whether the path space Markov property still holds, revealing the boundary of the framework's validity.

Watch

Extended reading notes

Core claim

Theorem 5.3 is the center: fix a boundary path b and condition the joint path law on it. If the conditional law of (Y,X) given B=b is dominated by a product reference Q^b_Y⊗Q^b_X, then B[0,T] is a Markov boundary between Y[0,T] and X[0,T] iff, for almost every b, the density H_b(y,x) factorizes as U_b(y)V_b(x). This turns a global statement about independent histories into a local factorization test. The paper then derives equivalent forms: in Itô diffusions the Girsanov log-likelihood of the boundary has no irreducible mixed term; in conditionally Markov clamped models the generator splits; in Gaussian models the (Y,X) precision block vanishes; at equilibrium the potential splits additively

Load-bearing premise

For almost every observed boundary path, the conditional distribution of the exterior and interior histories must be absolutely continuous with respect to some product of clamped reference distributions; if the conditional law is singular to every such product, the likelihood-ratio test that powers the entire framework is undefined.

Editorial extensions

If this is right

If a boundary path is a path space Markov boundary, the exterior history adds no predictive information about the interior beyond the boundary history: the optimal estimator of the interior from both histories is the estimator from the boundary alone.
Path-space screening implies instantaneous screening at each time, but the converse fails; conditioning on the full boundary path can remove dependence that reappears when averaging over boundary histories with the same instantaneous value.
In Gaussian path models, the Markov boundary condition reduces to a checkable algebraic test: the (Y,X) block of the path-space precision operator must vanish.
Under an isothermal convention, the minimum additional work needed to maintain a factorised (separated) architecture is exactly k_B T times the conditional mutual information between exterior and interior histories given the boundary history.
A vanishing boundary distinguishability does not imply zero entropy production: a nonequilibrium steady state can have an exact path space boundary while still dissipating housekeeping heat.

Reading between the lines

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

The framework suggests a practical diagnostic for model reduction: compute the boundary-conditioned log-likelihood of a candidate boundary path and test whether any irreducible mixed term between exterior and interior coordinates remains; the magnitude of that term directly prices the cost of ignoring the true coupling.
The mixture obstruction between path-level and time-level screening implies that standard instantaneous conditional-independence tests can miss substantial hidden path-level coupling; reliable screening claims require checking likelihood separation over entire histories, not just individual time slices.
In stochastic thermodynamics of molecular machines, one could test the criterion by measuring a motor coordinate as the boundary, computing the conditional likelihood of the external load history and internal state history, and checking whether the cross term vanishes; the predicted excess heat of enforcing factorization should then match observed dissipation.
The domination assumption is the main practical limitation: systems with deterministic constraints or singular couplings between exterior and interior may not admit any clamped product reference, so the factorization test needs a generalized formulation based on Lebesgue decomposition rather than a single Radon–Nikodym density.

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

Colleague,

The paper is a competent synthesis of path-space conditional independence, Girsanov theory, and information-theoretic thermodynamics. The core theorem (5.3) is just the Radon-Nikodym separability lemma applied to conditional laws; the paper is honest that the equivalences in the Gaussian and generator cases are standard. What is genuinely new is the packaging: the factorisation gap as excess free energy in §12.6 (Theorem 12.17, Corollary 12.18) and the maintenance bounds. These are simple corollaries of standard relative entropy identities, but I don't know another place where they are assembled this way, and they could be useful to people thinking about the thermodynamic cost of modular architecture. The measure-theoretic sections (2–5, 11) are correct and carefully stated.

The soft spots are real but fixable. Theorem 12.20 has a concrete metric error: since the control enters as M u in (12.15), the Girsanov relative entropy should be (1/(4 k_B T_bath)) E∫ ||u||^2_M dt, not ||u||^2_{M^{-1}}. Equations (12.17)–(12.18) inherit this. Theorem 12.9 also needs mutual absolute continuity for the exponential fluctuation identity; one-sided P_T << P†_T is not enough.

The bigger structural weakness is the domination hypothesis in Section 6.1. The abstract and Theorem 5.3 present log-likelihood separation as the path-space boundary characterization, but it only applies when the conditional law P^b is dominated by a clamped product reference Q^b_Y ⊗ Q^b_X. The paper never gives sufficient conditions in terms of the original SDE coefficients, and the assumption genuinely fails for degenerate boundary noise—e.g., dB_t = (Y_t+X_t) dt with σ_B = 0. Then conditioning on B=b forces Y and X onto a codimension-one set, the conditional law is singular relative to any nondegenerate product reference, and the density H_b does not exist. So the headline claim overstates scope. This is not a fatal flaw in the framework, but it should be stated as a limitation in the abstract, not buried in an assumption.

Who is this for? People at the intersection of stochastic thermodynamics and information theory—boundaries, modularity, information engines. It is a useful map of known results with a couple of new inequalities. The errors are not load-bearing for the core factorization theorems, but they are load-bearing for the specific thermodynamic bounds.

Recommendation: send to peer review, but require the Girsanov norm correction, the mutual absolute continuity fix, and an explicit discussion of the domination hypothesis and its failure modes. The paper deserves referee time.

Referee Report

3 major / 6 minor

Summary. The paper develops a unified framework for path-space Markov boundaries in stochastic systems decomposed into exterior (Y), boundary (B), and interior (X) variables. It defines the boundary property as conditional independence of the full paths Y_[0,T] and X_[0,T] given B_[0,T], and shows that, under domination by clamped product reference laws, this is equivalent to multiplicative separability of the conditional Radon–Nikodym density or additive separability of the boundary-path log-likelihood. For Itô diffusions the likelihood is computed via Girsanov, linking the factorisation gap to control energy and the Föllmer entropy identity. The paper also develops Gaussian precision-zero criteria, generator-splitting criteria, random-dynamical-system cocycle conditions, and thermodynamic interpretations: conditional mutual information as a distinguishability gap, a Girsanov maintenance bound for the cost of enforcing factorisation, and a decomposition of entropy production into excess and housekeeping parts. Numerous examples and convergence theorems are included.

Significance. If the technical results are corrected, the paper offers a valuable unification of conditional independence, path-likelihood inference, stochastic control, and stochastic thermodynamics. The core Radon–Nikodym separability criterion (Theorem 5.3) is clean and correct under its stated domination hypothesis. The Gaussian precision-zero and generator-splitting criteria are standard but gain a unified path-space presentation. The link between the factorisation gap and thermodynamic free-energy excess is attractive and potentially impactful. However, two central thermodynamic theorems contain specific technical errors that must be fixed before the quantitative claims can be relied upon. The paper's scope is also explicitly conditional on a domination assumption, and the manuscript would benefit from stating when this assumption holds and when it fails.

major comments (3)

[§12.6, Theorem 12.20, Eqs. (12.15)–(12.18)] The Girsanov metric in (12.16) is incorrect. In the controlled SDE (12.15) the drift difference between the controlled and uncontrolled laws is M u_t. With diffusion coefficient sigma = (2 k_B T M)^{1/2}, the squared Cameron–Martin norm of this drift difference is (M u)^T (2 k_B T M)^{-1} (M u) = (1/(2 k_B T)) u^T M u. The relative entropy is therefore H(P^{b,u}|P^b) = (1/(4 k_B T)) E^{P^{b,u}} ∫ ||u||^2_M dt, not ||u||^2_{M^{-1}} as written. The bounds (12.17) and (12.18) inherit this error. Please correct the metric and re-derive the maintenance bounds.
[§12.2, Theorem 12.9] The theorem assumes P_T ≪ P^†_T, but the proof uses the derivative dP^†_T/dP_T in the identity E_{P_T} exp(-Σ_T) = ∫ (dP^†_T/dP_T) dP_T. This derivative exists only if P^†_T ≪ P_T. Under the stated one-sided absolute continuity, the integral fluctuation identity may fail if P^†_T has mass on the zero set of P_T. The theorem should either assume mutual absolute continuity (P_T ≈ P^†_T) or the proof must be revised to handle the singular part. This affects the second-law statement and the entropy-production-rate result.
[§5, Theorem 5.3; §6.1, Theorem 6.1] The central necessary-and-sufficient characterisation is conditional on the domination assumption P^b ≪ Q^b_Y ⊗ Q^b_X for P_B-almost every b. The paper does not give sufficient conditions on the Itô coefficients for this domination to hold, and the condition fails in natural degenerate/noiseless boundary models. While the theorems are correctly stated with their hypotheses, the manuscript should explicitly delimit the domain of applicability: the log-likelihood separation criterion is not a universal characterization of path-space Markov boundaries, but only one under a non-trivial domination condition. A remark with concrete examples (e.g., uniformly elliptic diffusion vs. deterministic boundary dynamics) would prevent overgeneralisation of the central claim.

minor comments (6)

[§6.1] The phrase 'Brownian innovation boundary model' is informal. Please state precisely what independence of the exterior and interior innovation noises after conditioning on the boundary path means, and how it relates to the driving Brownian motions in (6.1).
[§12.1, Theorem 12.11] The assertion that the time-reversed stationary drift is b - 2v_π needs a derivation or reference, especially for nonconstant diffusion tensors D. The formula is standard in divergence form but should be justified for the reader.
[§12.6, Eq. (12.16)] The notation ||u||^2_{M^{-1}} presumes M is invertible. If M is only positive semidefinite, the inverse should be the Moore–Penrose pseudoinverse and the formula should be stated accordingly, as is done elsewhere in §12.3.
[§12.6, Example 12.23] In (12.21), the symbol C^b is used for both the full covariance matrix and the canonical-correlation matrix. Please distinguish these, e.g. by writing R^b for the latter, to avoid confusion.
[§6.2, Proposition 6.2] The sufficiency claim at the end ('In particular, if <r_Y, r_X> = 0 for all (y,x,b), then the likelihood separates') is valid, but the 'if and only if' in the first sentence requires the mixed term to separate as a functional of y plus a functional of x, which is trivially true for the zero function. Consider rewording to avoid circularity.
[General] The paper is very long and covers many subtopics. A table of notation and a short 'main theorems' summary would help the reader navigate the 18 sections.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the central equivalences are direct applications of the Radon–Nikodym separability lemma to definitions; no fitted inputs, no self-citations, and no prediction that reduces to its own input.

full rationale

The paper's derivation chain is self-contained. Theorem 5.3 is explicitly a direct application of Lemma 2.6 to the conditional law P^b: Lemma 2.6 proves that a dominated probability measure is a product measure if and only if its Radon–Nikodym density separates multiplicatively, and Definition 5.2 defines a path space Markov boundary as the factorization of the conditional law. Thus Theorem 5.3 is a legitimate mathematical equivalence, not a circular reduction. The same holds for Theorem 6.1 and Corollary 5.4, which are restatements of the same lemma under the explicitly stated Bayes–Girsanov density form (6.2). The thermodynamic results (Theorems 6.3, 12.11, 12.17, 12.20) are derived from standard external results—Girsanov's theorem, relative entropy identities, and the Fokker–Planck current algebra—not from fitted parameters or from the paper's own conclusions. There are no self-citations in the reference list; the only cited works by Nicoletti and Busiello are external sources for mutual information in changing environments. No parameter is fitted and no data are predicted. The domination hypothesis in Section 6.1 ('the theorem only needs domination by such product references') is an explicit scope condition: if it fails, as for noiseless deterministic boundaries, Theorems 5.3 and 6.1 simply do not apply. That is a generality limitation, not circularity. Accordingly, no circular step is identified and the circularity score is 0.

Assumptions & free parameters 0 free parameters · 9 assumptions · 0 invented entities

The paper introduces no free parameters or invented entities. It relies on standard measure-theoretic results, on the domain-specific assumption of domination by clamped product laws, and on several paper-specific structural conditions (cocycle boundary-respecting, initial factorisation, and an incorrect metric assumption in Theorem 12.20).

assumptions (9)

standard math Existence of regular conditional probabilities on Polish path spaces (standard Borel disintegration theorem).
Used throughout Section 2 and Section 3 to define conditional path laws; well-established.
domain assumption Conditional law Π^b is dominated by a clamped product law Q^b_Y ⊗ Q^b_X.
Assumed in Theorem 5.3 and 6.1; not automatic for path conditioning, and is load-bearing for the RN separability criterion.
domain assumption Weak well-posedness of the Itô system (6.1) and Novikov's condition for Girsanov densities.
Section 6 uses these to justify the Bayes-Girsanov form (6.2) and likelihood (6.6).
domain assumption Uniqueness of the martingale problem for the conditioned generator L^b_t.
Theorem 6.4 relies on uniqueness to conclude Π^b is the product law when the generator splits.
domain assumption Initial conditional factorisation Law(Y0,X0|S0,A0)=Law(Y0|S0,A0)⊗Law(X0|S0,A0) in the sensor-actuator system.
Theorem 7.2 assumes this; Lemma 15.1 proves a similar condition is necessary for the path-space property.
ad hoc to paper The cocycle respects the boundary: no common residual variable outside B remains after conditioning.
Introduced in Theorem 9.6 as an extra structural condition to make pushforwards preserve product structure.
ad hoc to paper For Theorem 12.9 the paper assumes only P_T ≪ P†_T, but the proof needs mutual absolute continuity so that dP†_T/dP_T exists.
The statement 'E exp(-Σ)=1' under one-sided AC is not valid as proved; in diffusion applications equivalence usually holds.
domain assumption For Theorem 12.11 and 12.13, v_π ∈ Range D and finite energy integrals / vanishing boundary terms in integration by parts.
Needed for finite entropy production and the housekeeping-excess decomposition.
ad hoc to paper Implicit metric choice in Theorem 12.20 that the control drift's Cameron-Martin norm is u^T M^{-1} u.
Eq (12.15) has drift M u, so the correct norm is u^T M u; the paper's use of M^{-1} makes (12.16) wrong.

how reviews work

0 comments

Cite this review

Pith. "Pith review of The nonequilibrium statistical mechanics of Markov interacting particles." pith.science (2026). https://pith.science/paper/2JEDQTLC

@misc{pith2026260713391,
  author       = {Pith},
  title        = {Pith review of: The nonequilibrium statistical mechanics of Markov interacting particles},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2JEDQTLC}},
  note         = {Machine review of arXiv:2607.13391}
}

read the original abstract

We consider coupled stochastic systems decomposed into exterior, boundary, and interior variables, with the boundary variables sometimes carrying the directed structure of a sensor and actuator. The central question is when the conditional law of histories factorises, and how this path space statement is detected by log likelihoods, by Girsanov changes of measure, and by information theoretic quantities used in nonequilibrium statistical physics. The basic object is a regular conditional probability on a path space. Under domination by clamped reference laws, the boundary property becomes multiplicative separation of a Radon--Nikodym derivative, or equivalently additive separation of a path log likelihood. For It\=o diffusions this log likelihood is computed by Girsanov's theorem; its expectation is the quadratic control energy appearing in the F\"ollmer entropy identity and in the stochastic control formulation of Schr\"odinger bridge problems. When exact factorisation fails, the remaining coupling is measured by conditional mutual information, namely the relative entropy between the true boundary-conditioned path law and the product of its conditional marginals. This gives a common language for boundary screening, path likelihood inference, controlled changes of path law, and the thermodynamic value of mutual information.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 1 linked inside Pith

[1]

Physical Review , volume =

Onsager, Lars and Machlup, Stefan , title =. Physical Review , volume =. 1953 , doi =

1953
[2]

Journal of Multivariate Analysis , volume =

Dembo, Amir and Zeitouni, Ofer , title =. Journal of Multivariate Analysis , volume =. 1991 , doi =

1991
[3]

, title =

Stuart, Andrew M. , title =. Acta Numerica , volume =. 2010 , doi =

2010
[4]

Dashti, Masoumeh and Law, Kody J. H. and Stuart, Andrew M. and Voss, Jochen , title =. Inverse Problems , volume =. 2013 , doi =

2013
[5]

Ayanbayev, Birzhan and Klebanov, Ilja and Lie, Han Cheng and Sullivan, T. J. , title =. arXiv preprint arXiv:2108.04597 , year =. 2108.04597 , archivePrefix =

arXiv
[6]

and Savage, Van M

Chang, Joshua C. and Savage, Van M. and Chou, Tom , title =. Journal of Statistical Physics , volume =. 2014 , doi =

2014
[7]

, title =

Girsanov, Igor V. , title =. Theory of Probability and Its Applications , volume =. 1960 , doi =

1960
[8]

An Entropy Approach to the Time Reversal of Diffusion Processes , booktitle =

F. An Entropy Approach to the Time Reversal of Diffusion Processes , booktitle =. 1985 , doi =

1985

Show all 27 references

[9]

A Variational Representation for Certain Functionals of Brownian Motion , journal =

Bou. A Variational Representation for Certain Functionals of Brownian Motion , journal =. 1998 , doi =

1998
[10]

Annales de l'Institut Henri Poincar

Lehec, Joseph , title =. Annales de l'Institut Henri Poincar. 2013 , doi =

2013
[11]

A Survey of the Schr

L. A Survey of the Schr. Discrete and Continuous Dynamical Systems , volume =. 2014 , doi =

2014
[12]

and Pavon, Michele , title =

Chen, Yongxin and Georgiou, Tryphon T. and Pavon, Michele , title =. Journal of Optimization Theory and Applications , volume =. 2016 , doi =

2016
[13]

and Bell, Anthony J

Still, Susanne and Sivak, David A. and Bell, Anthony J. and Crooks, Gavin E. , title =. Physical Review Letters , volume =. 2012 , doi =

2012
[14]

and Esposito, Massimiliano , title =

Horowitz, Jordan M. and Esposito, Massimiliano , title =. Physical Review X , volume =. 2014 , doi =

2014
[15]

Parrondo, Juan M. R. and Horowitz, Jordan M. and Sagawa, Takahiro , title =. Nature Physics , volume =. 2015 , doi =

2015
[16]

Physical Review Letters , volume =

Nicoletti, Giorgio and Busiello, Daniel Maria , title =. Physical Review Letters , volume =. 2021 , doi =

2021
[17]

Physical Review E , volume =

Nicoletti, Giorgio and Busiello, Daniel Maria , title =. Physical Review E , volume =. 2022 , doi =

2022
[18]

Physical Review X , volume =

Nicoletti, Giorgio and Busiello, Daniel Maria , title =. Physical Review X , volume =. 2024 , doi =

2024
[19]

1998 , doi =

Arnold, Ludwig , title =. 1998 , doi =

1998
[20]

, title =

Bogachev, Vladimir I. , title =
[21]

Da Prato, Giuseppe and Zabczyk, Jerzy , title =
[22]

2010 , edition =

Dembo, Amir and Zeitouni, Ofer , title =. 2010 , edition =

2010
[23]

and Wentzell, Alexander D

Freidlin, Mark I. and Wentzell, Alexander D. , title =. 2012 , edition =

2012
[24]

, title =

Pavliotis, Grigorios A. , title =
[25]

Physical Review Letters , volume =

Hatano, Takahiro and Sasa, Shin-ichi , title =. Physical Review Letters , volume =. 2001 , doi =

2001
[26]

Physical Review Letters , volume =

Esposito, Massimiliano and Van den Broeck, Christian , title =. Physical Review Letters , volume =. 2010 , doi =

2010
[27]

Reports on Progress in Physics , volume =

Seifert, Udo , title =. Reports on Progress in Physics , volume =. 2012 , doi =

2012

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [1]

Physical Review , volume =

Onsager, Lars and Machlup, Stefan , title =. Physical Review , volume =. 1953 , doi =

1953

[2] [2]

Journal of Multivariate Analysis , volume =

Dembo, Amir and Zeitouni, Ofer , title =. Journal of Multivariate Analysis , volume =. 1991 , doi =

1991

[3] [3]

, title =

Stuart, Andrew M. , title =. Acta Numerica , volume =. 2010 , doi =

2010

[4] [4]

Dashti, Masoumeh and Law, Kody J. H. and Stuart, Andrew M. and Voss, Jochen , title =. Inverse Problems , volume =. 2013 , doi =

2013

[5] [5]

Ayanbayev, Birzhan and Klebanov, Ilja and Lie, Han Cheng and Sullivan, T. J. , title =. arXiv preprint arXiv:2108.04597 , year =. 2108.04597 , archivePrefix =

arXiv

[6] [6]

and Savage, Van M

Chang, Joshua C. and Savage, Van M. and Chou, Tom , title =. Journal of Statistical Physics , volume =. 2014 , doi =

2014

[7] [7]

, title =

Girsanov, Igor V. , title =. Theory of Probability and Its Applications , volume =. 1960 , doi =

1960

[8] [8]

An Entropy Approach to the Time Reversal of Diffusion Processes , booktitle =

F. An Entropy Approach to the Time Reversal of Diffusion Processes , booktitle =. 1985 , doi =

1985