Mean-field limits \`a la Tanaka and large deviations for particle systems with network interactions

Antoine Diez; Louis-Pierre Chaintron

REVIEW 2 major objections 3 minor 8 references

Mean-field limits \`a la Tanaka and large deviations for particle systems with network interactions

T0 review · 2 major / 3 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 This paper proves that non-exchangeable particle systems with network interactions—constant or adaptive—can be analyzed through one fixed-point scheme, yielding a mean-field limit and a large-deviation principle whose rate function is a rel

desk verdict Solid Tanaka-framework paper, but Theorem 4.4 overreaches: the LDP for adaptive networks rests on an identity that only holds for constant weights. read the letter →

arxiv 2510.04894 v2 pith:6FWXABFC submitted 2025-10-06 math.PR math.AP

Louis-Pierre Chaintron , Antoine Diez This is my paper · ORCID

classification math.PRmath.AP MSC 60F1060K3535Q84

keywords mean-fieldlimitnon-exchangeableparticlesystemsdigraphmeasureslargedeviationsinteractionadaptivenetworksrelativeentropyratefunctionnetworkinteractions

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 aims to show that many non-exchangeable interacting particle systems—where each particle interacts through its own weighted network of connections rather than through a common empirical measure—fit into one abstract fixed-point scheme. In that scheme, each particle is the output of a map applied to its own input data (label, initial state, noise), while the whole network is encoded as a digraph measure assigning to each label a measure on weight–label pairs. The paper proves that if this network encoding converges and is Lipschitz with respect to labels, then the particle system converges to a mean-field limit, and the laws of the interaction measures satisfy a large-deviation principle with a rate function given explicitly by relative entropy. It also gives a formal PDE characterization of the limit for adaptive networks in an important linear case. A sympathetic reader should care because this provides a unified route from graph-structured microscopic models to macroscopic statistical descriptions, with quantitative fluctuation information attached.

What carries the argument

The machinery is the fixed-point map X^α defined by equation (5), parameterized by α ∈ A. For networks, A consists of Lipschitz maps from the input space Ω to probability measures on [0,1]×Ω, so that α(ω) encodes the outgoing edge weights of the label in ω; this is the digraph-measure formalism, i.e. a map sending each vertex label to the empirical measure of its edge weights and target labels. The interaction measure π_t(α,ω,X)=(W_t(α,ω,X),X)_# α(ω) assembles weights and partner trajectories, and the drift averages this measure against a kernel K. The decisive mechanism is the Lipschitz dependence of X^α on (α,ω) proved in Theorem 3.1, which turns metric convergence of the digraph measures

What would settle it

Take a weight matrix w^N_{ij}=W(i/N,j/N) with W(ξ,ξ′) Lipschitz in the first label uniformly in the second but not jointly Lipschitz, satisfying the paper's convergence conditions, simulate the particle system (10), and check whether the empirical interaction measures converge in bounded-Lipschitz distance to the predicted minimizer and whether their exponential rate is K(π)=H(π|(Id,Y^π)_# α(ξ,0,0)); any passage through all hypotheses with a different rate or a failed concentration would contradict the central theorem.

Watch

Extended reading notes

Core claim

The central claim is Theorem 4.4: for each label ξ, the empirical interaction measures π_N(ξ)—the joint distribution of edge weights and partner trajectories seen by vertex ξ—obey a large-deviation principle with good rate function K(π)=H(π | (Id,Y^π)_# α(ξ,0,0)), where Y^π solves the fixed-point equation (20) and H is relative entropy. Underlying this is an abstract result (Theorem 3.1 and its corollaries): whenever a particle system is realized as a fixed point of a Lipschitz map of input data and parameters, convergence of the parameter and input measures propagates to convergence of the particles, and any large-deviation principle on the inputs contracts to one on the empirical measures.

Load-bearing premise

The load-bearing premise is that the network's digraph measures are uniformly continuous in the label and converge to a deterministic limit with a Laplace-type condition; sparse random graphs with vanishing connection probabilities violate this and fall outside the theorems.

Editorial extensions

If this is right

For any network satisfying the label-continuity and convergence conditions, the interaction measures converge almost surely to the unique minimizer of the rate function, which is also the fixed point of the map π ↦ (Id,Y^π)_# α(ξ).
The rate function's relative-entropy form makes rare-event analysis quantitative: the cost of observing an interaction measure π is the entropy of π against the reference law built from Y^π, computable by solving the fixed-point equation (20).
The mean-field limit extends beyond constant networks to adaptive ones, covering systems where edge weights evolve by an ODE depending on both endpoint trajectories; in the linear case the limit law is characterized by the coupled PDE system of Corollary 5.3.
The abstract fixed-point theorem gives pathwise convergence of labeled particles with rates depending on the distance between the discrete digraph measure and its limit, so quantitative convergence bounds follow in this setting.
The framework covers environment-noise models and dense random graphs, where the digraph limit is an average over an independent random environment; in these cases the large-deviation principle follows from the labeled Laplace-condition theorem proved in the appendix.

Reading between the lines

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

The same contraction argument should yield a central limit theorem for interaction measures whenever the input digraph measures satisfy a CLT and the fixed-point map is differentiable in the parameter; the paper only remarks on this possibility in the exchangeable setting.
The rate-function expression suggests a practical diagnostic for graph-coupled particle data: compute the auxiliary fixed point Y^π and compare the relative entropy of observed interaction measures against the predicted reference law to detect rare or misspecified network configurations.
The label-Lipschitz requirement could plausibly be relaxed in a two-scale direction: if digraph measures converge in a weaker topology but the input empirical measures are regularized, a mean-field limit may still hold even where the large-deviation statement becomes unclear.
A natural test case outside the paper is multiplicative or path-dependent noise: since the fixed-point contraction only needs Lipschitz dependence on the noise path, augmenting the input space with a second noise component may yield analogous large-deviation results without changing the rate-function structure.

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

First thing you should know: this is a good paper with one overreach. The Tanaka fixed-point framework is clean, and the abstract results (Theorem 3.1, Corollaries 3.2-3.3) are solid. The mean-field limit for non-exchangeable systems via digraph measures is a genuine extension of [Cog+20], and the labeled Sanov theorem in Appendix B is a useful tool in its own right. The main new result, Theorem 4.4, gives an LDP for the interaction measure with a relative-entropy rate function. That is a nice addition for constant-weight networks like Erdős-Rényi or Lipschitz graphons.

But the theorem is stated for the general adaptive network system (10), and that's where it breaks. The proof relies on identity (21), which says X^α equals Y^{(Id,X^α)_#α(ω)}(ω). That identity holds when the weights are constant (W_t ≡ w'), because then X^α and the Y^π fixed point solve the same equation. Once the weights evolve through (9), the drift of X^α contains W_t(X^α(ω), X^α(ω'), w'), while the Y^π equation uses the frozen initial weight w'. These are different maps, and there is no reason for their fixed points to agree. So (21) fails, and the contraction-principle computation in the proof of Theorem 4.4 does not go through for adaptive weights. The theorem should be restricted to constant weights, or the rate function has to be re-derived. This is not a purely cosmetic issue: the abstract and introduction advertise adaptive networks as a main feature, and Section 5 (which is explicitly informal) depends on the adaptive dynamics. The claimed LDP for adaptive networks is currently unsupported.

Everything else is in better shape. The abstract well-posedness, the Lipschitz estimates, the mean-field limit, and the labeled Sanov theorem check out. The PDE section is honestly labeled as formal, with the smoothness assumptions flagged. The citation pattern looks fair, with appropriate credit to [Cog+20], [KX22], [KP24], and others.

Bottom line: the paper deserves a serious referee, but the referee should ask for a fix to Theorem 4.4. If the authors restrict it to constant weights, the paper stands as a solid contribution. For adaptive weights, they need either a counterexample or a genuinely new argument.

Referee Report

2 major / 3 minor

Summary. The paper proposes an abstract fixed-point framework in the spirit of Tanaka to study non-exchangeable mean-field particle systems. A parameter α in an abstract metric space encodes the interaction structure; particles are realised as X^α(ω_i), where X^α solves a fixed point. Under boundedness and Lipschitz conditions, the authors prove well-posedness, continuity in parameters, mean-field limits, and an LDP via the contraction principle. They specialise to digraph measures to model network interactions, including time-evolving weights, and derive a mean-field limit and an LDP for the interaction measure with a relative-entropy rate function. A final section gives a formal PDE closure for the limit in some constant and adaptive network cases.

Significance. The abstract results (Theorem 3.1, Corollaries 3.2–3.3, Lemma 4.2, Theorem B.1) are rigorous and elegant; the labeled Sanov theorem is a useful standalone contribution, and the relative-entropy form of the rate function is appealing. However, the main LDP for adaptive networks (Theorem 4.4) is not established as stated because its proof relies on an identity that only holds for constant weights. This is a load-bearing flaw in a central advertised result. The mean-field limit and the constant-weight LDP are valuable, but the paper overreaches in claiming adaptive networks in the LDP.

major comments (2)

[§4.2, Eq. (21) and Theorem 4.4] Identity (21) is false for adaptive weights. In the particle system (10) with weight dynamics (9), the drift of X^α involves W_t(X^α(ω), X^α(ω'), w'), the time-evolved weight obtained from the ODE. The auxiliary process Y^π defined in (20) uses the frozen initial weight w' as its interaction coefficient. These are different fixed-point problems; they coincide only when φ ≡ 0 (constant weights). The proof of Theorem 4.4 explicitly uses (21) to simplify the contraction-principle rate function to K(π)=H(π|(Id,Y^π)_#α(ξ,0,0)), so the stated LDP for the general adaptive system is unjustified. The theorem overclaims: it is at most valid for constant networks, and the true rate function for adaptive networks would involve the weight-evolution map, not the one stated.
[§4.2, condition (28) and examples] The LDP in Theorem 4.4 is conditional on the Laplace-principle condition (28). The paper claims that (28) holds for Erdős-Rényi graphs and random-environment models, but the verification is only a remark stating that it 'reduces to the Riemann sum convergence theorem'. No detailed proof is supplied. Since the examples are part of the advertised applications, the authors should either provide a complete verification of (28) for each claimed example or state the theorem as purely conditional without claiming those applications.

minor comments (3)

[§5.2, Proposition 5.1] Typo: 'we de not specify' should be 'we do not specify'. Also, the informal nature of Proposition 5.1 and Corollary 5.3 should be clearly flagged in the abstract or introduction, as the current wording 'closed PDE characterization' may be read as a rigorous result.
[General] The title uses '`a la Tanaka'; please use the proper accent 'à la Tanaka'. In the proof of Lemma 4.2, the notation P^N is introduced but it is not used consistently; clarify its role.
[Remark 4.6] The phrase 'could be written more concisely' is unclear; perhaps mean 'could be treated more directly'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper’s results follow from explicit regularity, graph-limit and Laplace-principle hypotheses via fixed-point, Sanov, Gärtner–Ellis and contraction arguments; self-citations are not load-bearing.

full rationale

The derivation chain is self-contained in the sense required here. The abstract framework (Theorem 3.1) is a Banach fixed-point/Gronwall argument; Corollaries 3.2–3.3 are direct continuity and contraction-principle consequences. The network mean-field limit (Proposition 4.3) and the large deviation principle (Theorem 4.4) are obtained from explicit hypotheses on the digraph measures—uniform label-continuity (15), convergence (16), and the labeled Sanov Laplace condition (28)—and from Theorem B.1, which is proved in the appendix via standard Sanov/Gärtner–Ellis methods. The rate function K(π)=H(π|(Id,Y^π)_#α(ξ,0,0)) is a self-consistency/fixed-point description of the limiting interaction measure; this is the intended content of a McKean–Vlasov LDP rather than a tautology, since the large-deviation speed and the reference measure originate from the input digraph Sanov theorem, not from assuming the conclusion. No parameter is fitted to data and no empirical quantity is relabelled as a prediction. Section 5 is explicitly marked informal and does not feed back into Theorem 4.4. Self-citations ([Cha24], [CD22a], [CD22b]) are contextual reviews or side remarks, not load-bearing unique-ness imports. The principal caveat—that identity (21) appears to require constant weights W_t ≡ w′ and may fail for adaptive weights, so the stated rate function in Theorem 4.4 may be unjustified for adaptive networks—is a mathematical correctness concern, not a circularity of the kind defined here. Likewise, the admitted restrictions on sparse Erdős–Rényi graphs (Remark 4.6) and the unproved smoothness assumption in Proposition 5.1 are scope/rigor limitations, not circularities.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The central results are conditional on standard stochastic-analysis assumptions and on explicit regularity/convergence conditions for the network structure; no parameters are fitted to data and no new physical entities are introduced. The formal PDE section adds extra smoothness and structural assumptions that are not justified.

assumptions (5)

domain assumption Lipschitz regularity and global boundedness of the interaction kernel K and the weight-evolution function φ.
Used throughout to apply the Banach fixed-point theorem and Gronwall estimates (Theorem 3.1, Lemma 4.1); assumed uniformly in time.
domain assumption Label-continuity and convergence of digraph measures (conditions (15) and (16)).
Assumed to ensure the empirical digraph measures α_N converge to α uniformly in the label ξ; a key restriction excluding non-continuous graph limits.
domain assumption Laplace principle / LDP condition (28) for the input digraph measures.
Theorem 4.4 and Theorem B.1 require the exponential integrals in (28) to converge to Λ(φ); verified for i.i.d. weights and Lipschitz graphons but not for sparse random graphs (Remark 4.6).
standard math Standard large-deviation toolkit: Sanov, Gärtner–Ellis, contraction principle, marginal disintegration, entropy contraction.
Used in Corollary 3.3 and Appendices A/B; textbook results from Dembo–Zeitouni and Fischer.
ad hoc to paper Smooth positive density H_t and the decoupled linear weight dynamics φ=ℓ+αw in Section 5.
Proposition 5.1 explicitly assumes these to close the PDE system; existence of H_t is not proven and the proposition is labeled informal.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Mean-field limits \`a la Tanaka and large deviations for particle systems with network interactions." pith.science (2026). https://pith.science/paper/6FWXABFC

@misc{pith2026251004894,
  author       = {Pith},
  title        = {Pith review of: Mean-field limits \`a la Tanaka and large deviations for particle systems with network interactions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6FWXABFC}},
  note         = {Machine review of arXiv:2510.04894}
}

read the original abstract

This article proposes a unified framework to study non-exchangeable mean-field particle systems with some general interaction mechanisms. The starting point is a fixed-point formulation of particle systems originally due to Tanaka that allows us to prove mean-field limit and large deviation results in an abstract setting. While it has been recently shown that such formulation encompasses a large class of exchangeable particle systems, we propose here a setting for the non-exchangeable case, including the case of adaptive interaction networks. We introduce sufficient conditions on the network structure that imply the mean-field limit and a new large deviations principle for the interaction measure. Finally, we formally highlight important models for which it is possible to derive a closed PDE characterization of the limit.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

8 extracted references · 5 linked inside Pith

[1]

Large-Population Limits of Non-exchangeable Particle Systems

[AD24] N. Ayi and N. P. Duteil. “Large-Population Limits of Non-exchangeable Particle Systems”. In: Active Particles, Volume 4 . Ed. by J. A. Carrillo and E. Tadmor. Cham: Springer Nature Switzerland, 2024, pp. 79–133. [Bac+20] J. Backhoff, G. Conforti, I. Gentil, and C. L´ eonard. “The Mean Field Schr¨ odinger Problem: Ergodic Behavior, Entropy Estimates...

arXiv 2024
[3]

The Mean Field Equa- tion for the Kuramoto Model on Graph Sequences with Non-Lipschitz Limit

University of California Press Berkeley and Los Angeles, California, 1956, pp. 171–197. [KM18] D. Kaliuzhnyi-Verbovetskyi and G. S. Medvedev. “The Mean Field Equa- tion for the Kuramoto Model on Graph Sequences with Non-Lipschitz Limit”. In: SIAM J. Math. Anal. 50.3 (2018), pp. 2441–2465. [KP24] C. Kuehn and C. Pulido. “Mean-Field Limits for Stochastic In...

arXiv 1956
[19]

The Continuum Limit of the Kuramoto Model on Sparse Random Graphs

Van Nostrand Reinhold Company, 1969, pp. 177–194. [Med19] G. S. Medvedev. “The Continuum Limit of the Kuramoto Model on Sparse Random Graphs”. In: Commun. Math. Sci. 17.4 (2019), pp. 883–

1969
[155]

Pathwise McKean– Vlasov Theory with Additive Noise

[Cog+20] M. Coghi, J.-D. Deuschel, P. K. Friz, and M. Maurelli. “Pathwise McKean– Vlasov Theory with Additive Noise”. In: Ann. Appl. Probab. 30.5 (2020). 33 [CDG20] F. Coppini, H. Dietert, and G. Giacomin. “A Law of Large Numbers and Large Deviations for Interacting Diffusions on Erd˝ os–R´ enyi Graphs”. In: Stochastics and Dynamics 20.02 (2020), p. 20500...

arXiv 2020
[488]

A Glivenko-Cantelli Theorem for Empirical Measures of Independent but Non-Identically Distributed Random Variables

[Wel81] J. A. Wellner. “A Glivenko-Cantelli Theorem for Empirical Measures of Independent but Non-Identically Distributed Random Variables”. In: Stochastic Process. Appl. 11.3 (1981), pp. 309–312. 36

1981
[923]

Interacting Diffusions on Random Graphs with Diverging Average Degrees: Hydrodynamics and Large Deviations

35 [OR19] R. I. Oliveira and G. H. Reis. “Interacting Diffusions on Random Graphs with Diverging Average Degrees: Hydrodynamics and Large Deviations”. In: J. Stat. Phys. 176.5 (2019), pp. 1057–1087. [PT24] T. Paul and E. Tr´ elat. “From Microscopic to Macroscopic Scale Equa- tions: Mean Field, Hydrodynamic and Graph Limits”. In:arXiv preprint: arXiv:2209....

arXiv 2019
[1996]

A New Model for Self-Organized Dynamics and Its Flocking Behavior

[MT11] S. Motsch and E. Tadmor. “A New Model for Self-Organized Dynamics and Its Flocking Behavior”. In: J. Stat. Phys. 144.5 (2011), p

2011
[2009]

Recent Progress on Limit Theorems for Large Stochastic Particle Systems

[Fat+23] M. Fathi, P. Le Bris, A. Menegaki, P. Monmarche, J. Reygner, and M. Tomasevic. “Recent Progress on Limit Theorems for Large Stochastic Particle Systems”. In: ESAIM: Proceedings and Surveys 75 (2023). Ed. by M. Doumic, S. Gadat, and Q. M´ erigot, pp. 2–23. [Fis14] M. Fischer. “On the Form of the Large Deviation Rate Function for the Empirical Meas...

arXiv 2023

Pith tools

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

[1] [1]

Large-Population Limits of Non-exchangeable Particle Systems

[AD24] N. Ayi and N. P. Duteil. “Large-Population Limits of Non-exchangeable Particle Systems”. In: Active Particles, Volume 4 . Ed. by J. A. Carrillo and E. Tadmor. Cham: Springer Nature Switzerland, 2024, pp. 79–133. [Bac+20] J. Backhoff, G. Conforti, I. Gentil, and C. L´ eonard. “The Mean Field Schr¨ odinger Problem: Ergodic Behavior, Entropy Estimates...

arXiv 2024

[2] [3]

The Mean Field Equa- tion for the Kuramoto Model on Graph Sequences with Non-Lipschitz Limit

University of California Press Berkeley and Los Angeles, California, 1956, pp. 171–197. [KM18] D. Kaliuzhnyi-Verbovetskyi and G. S. Medvedev. “The Mean Field Equa- tion for the Kuramoto Model on Graph Sequences with Non-Lipschitz Limit”. In: SIAM J. Math. Anal. 50.3 (2018), pp. 2441–2465. [KP24] C. Kuehn and C. Pulido. “Mean-Field Limits for Stochastic In...

arXiv 1956

[3] [19]

The Continuum Limit of the Kuramoto Model on Sparse Random Graphs

Van Nostrand Reinhold Company, 1969, pp. 177–194. [Med19] G. S. Medvedev. “The Continuum Limit of the Kuramoto Model on Sparse Random Graphs”. In: Commun. Math. Sci. 17.4 (2019), pp. 883–

1969

[4] [155]

Pathwise McKean– Vlasov Theory with Additive Noise

[Cog+20] M. Coghi, J.-D. Deuschel, P. K. Friz, and M. Maurelli. “Pathwise McKean– Vlasov Theory with Additive Noise”. In: Ann. Appl. Probab. 30.5 (2020). 33 [CDG20] F. Coppini, H. Dietert, and G. Giacomin. “A Law of Large Numbers and Large Deviations for Interacting Diffusions on Erd˝ os–R´ enyi Graphs”. In: Stochastics and Dynamics 20.02 (2020), p. 20500...

arXiv 2020

[5] [488]

A Glivenko-Cantelli Theorem for Empirical Measures of Independent but Non-Identically Distributed Random Variables

[Wel81] J. A. Wellner. “A Glivenko-Cantelli Theorem for Empirical Measures of Independent but Non-Identically Distributed Random Variables”. In: Stochastic Process. Appl. 11.3 (1981), pp. 309–312. 36

1981

[6] [923]

Interacting Diffusions on Random Graphs with Diverging Average Degrees: Hydrodynamics and Large Deviations

35 [OR19] R. I. Oliveira and G. H. Reis. “Interacting Diffusions on Random Graphs with Diverging Average Degrees: Hydrodynamics and Large Deviations”. In: J. Stat. Phys. 176.5 (2019), pp. 1057–1087. [PT24] T. Paul and E. Tr´ elat. “From Microscopic to Macroscopic Scale Equa- tions: Mean Field, Hydrodynamic and Graph Limits”. In:arXiv preprint: arXiv:2209....

arXiv 2019

[7] [1996]

A New Model for Self-Organized Dynamics and Its Flocking Behavior

[MT11] S. Motsch and E. Tadmor. “A New Model for Self-Organized Dynamics and Its Flocking Behavior”. In: J. Stat. Phys. 144.5 (2011), p

2011

[8] [2009]

Recent Progress on Limit Theorems for Large Stochastic Particle Systems

[Fat+23] M. Fathi, P. Le Bris, A. Menegaki, P. Monmarche, J. Reygner, and M. Tomasevic. “Recent Progress on Limit Theorems for Large Stochastic Particle Systems”. In: ESAIM: Proceedings and Surveys 75 (2023). Ed. by M. Doumic, S. Gadat, and Q. M´ erigot, pp. 2–23. [Fis14] M. Fischer. “On the Form of the Large Deviation Rate Function for the Empirical Meas...

arXiv 2023