{"id":"e11018d8-211c-4a81-a286-a91b92d1a08f","arxiv_id":"2502.09506","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Stochastic RG flows on a finite volume with frozen empirical magnetization yield formal Fokker-Planck equations for the magnetization's large-deviation rate function, but no new rate function is computed.","lead":"This paper recasts Wilson's exact renormalization group as a Fokker-Planck diffusion and adapts Carosso's stochastic coarsening scheme to a finite volume with a frozen average field. It builds a formal machinery aimed at the large-deviation probability of the empirical magnetization, but stops short of computing any concrete rate function.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The constrained Carosso flow conserves p_L(m_e) by construction, so the advertised \"gradual integration\" does not actually derive or compute the large-deviation rate function; the noise amplitude D remains an unconstrained regulator parameter.","rationale":"I read the paper as a pedagogical reformulation rather than a new computational method. The Wilson-Kogut and Carosso exact-RG equations are internally consistent, and the Fokker-Planck/Langevin correspondence is correctly developed. The constrained construction of Section VII is mathematically sound: the marginal is conserved, and the flows for E_t and W_t follow from the Fokker-Planck equation. The concern is about the central claim in the abstract: the word 'obtained' suggests that the stochastic RG yields p_L(m_e) by gradual integration, but the construction preserves p_L(m_e) by design and therefore cannot produce it from the model unless it is already known. The rate function is the constrained minimum of the original E0 (Eq. 58); the Carosso time evolution leaves this minimum invariant. This is not a contradiction of the formal equations, but it undercuts the advertised application. The paper acknowledges the need to derive noise statistics and lists future directions, which supports a conditional rather than reject verdict. My check would expose the tautological character of the conservation by showing that an incorrect initial marginal is never corrected by the flow.","tokens_in":48191,"tokens_out":11614,"duration_ms":125657,"concrete_test":"Initialize the constrained Carosso SDE (Eqs. 142-143) for a non-Gaussian model (e.g., phi^4 in d=1 with a few Fourier modes) from an initial distribution that has the model's correct conditional distribution for nonzero modes but a deliberately wrong (e.g., Gaussian) marginal in me. Run to large t and record the empirical distribution of me. Because the transition kernel is independent of me and normalized, the me marginal is exactly preserved, so the output marginal is the input Gaussian, not the true phi^4 p_L(me). This demonstrates that the construction cannot obtain p_L(me) from the model's interactions: it must be supplied as an initial condition. If instead one initializes from the true P0, the procedure simply returns the input marginal, and the rate function still requires the original minimization in Eq. 58.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central claim is that p_L(m_e) can be obtained from gradual integration over nonzero-momentum Fourier coefficients via an adaptation of the Carosso stochastic RG. In Section VII this adaptation freezes m_e and lets the nonzero modes evolve as independent Ornstein-Uhlenbeck processes with noise from which the zero mode has been removed (Eqs. 134-143). The amplitude D is a free regulator parameter. Equation 153 then shows that p_L[me] is conserved by construction: the factorized OU propagator integrates to unity over the final nonzero modes for every t. This is not an integration over eliminated degrees of freedom; it is an invariant imposed on the stochastic process. The finite-time solution for the effective energy, Eq. 160, is an inf-convolution of E0 with a Gaussian kernel, and its minimum over the fluctuating modes is exactly min_{phi0} E0 (attainable by choosing phi = e^{-omega t} phi0). Hence the rate function i(me) in Eq. 58 is just the original constrained minimization of E0; the RG flow contributes no new information about p_L(me). The paper itself (Section VI D 2) defers the derivation of the noise statistics from the model's interactions, and the conclusion lists the application as future work. For the formal exact-RG flow the arbitrariness of D is acceptable as a regulator choice, but for the advertised large-deviation application it is load-bearing: nothing in the construction maps the chosen OU kernel to the actual marginalization of the phi^4-type model. If the initial marginal is not already known, the flow cannot produce it; if it is known, the flow merely returns it.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a pedagogical review of stochastic formulations of exact renormalization group flows: the Wilson-Kogut \"incomplete integration\" scheme as a linear Fokker-Planck flow for independent Ornstein-Uhlenbeck modes, the Wegner-Morris continuity-equation perspective, and the Carosso interpretation of the associated Langevin dynamics as genuine field coarsening. After this review, the paper adapts the Carosso RG to a finite volume L^d with periodic boundary conditions, freezes the empirical magnetization m_e, imposes a zero-mean constraint on the noise, and derives explicit Langevin/Fokker-Planck equations for the nonzero Fourier modes, an effective-energy flow (Eq. 157), a finite-time saddle-point solution (Eq. 160), and a generating-function flow (Eq. 163). The abstract and conclusion claim that p_L(m_e) can thereby be obtained by gradual integration over all nonzero-momentum Fourier coefficients, yielding the large-deviation rate function i(m_e).","tokens_in":48508,"tokens_out":4510,"duration_ms":43139,"significance":"The pedagogical part of the manuscript is clearly written and largely internally consistent; the explicit propagators, Fokker-Planck equations, and finite-time flow equations (Eqs. 144, 147, 157, 160, 163) are useful reference results and the connections among Wilson, Wegner-Morris, and Carosso formulations are presented accessibly. The advertised new application, however, is not delivered: in Section VII the distribution p_L(m_e) is conserved by construction rather than computed by a gradual integration, and the derivation of the noise statistics from the underlying interacting model is explicitly deferred. The paper is therefore best assessed as a valuable exposition and a proposal for a class of conditional stochastic RG flows, not as a demonstration that the Carosso RG computes the large-deviation properties of p_L(m_e).","major_comments":[{"comment":"The central claim of the abstract and of Section VIII is that p_L(m_e) can be obtained by \"gradual integration\" over the nonzero-momentum Fourier coefficients via the Carosso stochastic RG. In Section VII F, however, the conservation of p_L[m_e] is imposed by construction: the zero-momentum mode is frozen (Eq. 130), the noise is constrained to have zero spatial integral (Eq. 134), and each Ornstein-Uhlenbeck propagator integrates to unity over the final nonzero Fourier coefficients for every t, giving Eq. 153 as an identity. Consequently, the method does not compute p_L(m_e) or the rate function i(m_e); Eq. 58 remains the original constrained minimization of E0, and Eq. 161 reduces at t = +infinity to min_{phi0} E0[m_e; phi0]. The advertised incremental marginalization is never actually performed.","section":"Section VII F, Eqs. (130), (134), (153)"},{"comment":"The Gaussian white noise in the Carosso Langevin SDE is postulated, with amplitude D left as a free parameter, and the manuscript explicitly defers to future work the derivation of the noise statistics from the interactions of the eliminated modes. For a formal exact-RG flow this regulator freedom is acceptable, but for the claimed large-deviation application it is load-bearing: unless the noise amplitude and correlation structure are those generated by integrating out the nonzero modes of the phi^4-type model, the Carosso flow is an arbitrary stochastic deformation and the conserved p_L(m_e) is not shown to be the original model's marginal. Since Eq. 153 holds for any D, the parameter D carries no information about the original p_L(m_e); the manuscript should either derive D from the microscopic model or explicitly state that the construction defines a different conditional stochastic process.","section":"Section VI D 2, Eqs. (119), (135)"},{"comment":"The simplified RG flow for the intensive energy drops the O(1/L^d) terms present in Eq. 157, retaining only the leading-order contributions. This is a legitimate large-L saddle-point approximation, but the wording around Eq. 63 and the conclusion presents the procedure as an exact finite-volume RG. The manuscript should state clearly that Eq. 158 is a leading-order truncation and explain why the discarded O(1/L^d) terms do not affect the claimed relation between the flow and p_L(m_e). As written, the exactness of the finite-volume RG flow is overstated.","section":"Section VII G, Eq. (158)"}],"minor_comments":[{"comment":"The heading contains a typo: \"Orstein-Uhlenbeck\" should be \"Ornstein-Uhlenbeck.\"","section":"Section VII D heading"},{"comment":"The conclusion says the magnetization application was discussed in \"sections II, III and IV,\" but Section IV is the Wilson-Kogut review, not the finite-volume magnetization construction; this should be \"sections II, III and VII.\"","section":"Section VIII, paragraph (ii)"},{"comment":"The notation P_t[h; ...] with square brackets for a probability density over Fourier coefficients is nonstandard and may be confused with a functional or a generating function; a lowercase p_t would be clearer.","section":"Section VII F, Eq. (153)"},{"comment":"The paper cites the recent literature on large deviations of the magnetization [13-18] but does not compare the conditional-OU construction with the exact functional RG methods used there; a short comparison would help the reader judge the claimed novelty.","section":"Section III and VII"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a competent and useful review of stochastic formulations of exact RG, and the derivations in Section VII are internally consistent. The difficulty is that the central advertised result — computing p_L(m_e) by gradual integration — is reduced to a tautology by the constrained-noise construction, and the noise statistics are not derived from the model. A revision that reframes the paper as a pedagogical review with a carefully worded proposal, and that removes or substantially qualifies the large-deviation claim, could be publishable; in its current form the abstract and conclusion overstate what is shown."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe useful part of this paper is the first six sections: a genuinely careful, self-contained account of how Wilson's incomplete integration becomes a linear Fokker-Planck flow in OU modes, how Wegner-Morris fits in, and what Carosso's Langevin interpretation adds. The finite-volume adaptation in Section VII—freezing m_e, constraining the noise to zero mean, and writing explicit flows for the effective energy and generating function—is worked out in detail and the algebra is consistent. I checked the noise normalization (Eq. 140), the propagator (144), the energy flow (157), and the generating function flow (163); they hold together.\n\nThe soft spot is the advertised application. The abstract and goal sections say p_L(m_e) can be obtained from gradual integration over the non-zero modes. In Section VII it is not integrated out to obtain p_L(m_e); the zero-mean noise constraint (Eq. 134) and frozen m_e make p_L(m_e) conserved by construction (Eq. 153). The rate function that emerges at t=infty (Eq. 161) is just the original constrained minimization of E_0. So the construction preserves the marginal if you already know it; it does not compute or derive it for an interacting model. The paper is honest in Section VI D 2 that the noise statistics are assumed, not derived, and the conclusion lists the application as future work—but the abstract overstates what is achieved.\n\nThat said, the pedagogical content is real and the formal part is not circular in a damaging sense; it is a well-defined stochastic deformation with D as a free regulator. The gap is between what the flow can deliver and what the title/abstract promise. For the pedagogical review alone, this deserves a serious referee. I would recommend acceptance after the authors revise the claims: either drop the large-deviation application from the abstract or present it as a proposal with an explicit test (e.g., a non-Gaussian phi^4 case where p_L(m_e) is computed and compared with exact RG results). As it stands, I would not cite it for the large-deviation result, but I would cite it for the clear exposition of the Fokker-Planck/Carosso view.\n\nBottom line: worthwhile paper, honest math, overstated application. Send to a referee who knows exact RG and ask them to push on the distinction between an invariant and a computation.","headline":"Solid pedagogical review of stochastic RG; the new finite-volume scheme conserves, doesn't compute, p_L(m_e) despite the abstract's promise.","tokens_in":49059,"tokens_out":2956,"would_cite":true,"duration_ms":28428,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["82B28","60J60","60H15","82C31"],"pacs":["05.10.Gg","05.40.-a","64.60.-i"],"model":"deepseek-v4-flash","headline":"The paper recasts exact renormalization group flows as stochastic coarsening processes, preserving the distribution of the empirical magnetization by construction.","keywords":["exact renormalization group","Fokker-Planck equation","Ornstein-Uhlenbeck process","Langevin stochastic differential equation","stochastic coarsening","empirical magnetization","large deviations","Edwards-Wilkinson dynamics"],"falsifier":"Run the Carosso flow for a simple interacting theory such as $\\phi^4$ on a small finite-volume lattice and compare the exact marginal $p_L(m_e)$ with the marginal obtained by integrating the flow to $t=\\infty$ with the assumed white noise. If the two disagree, the noise statistics are not those generated by the eliminated modes and the stochastic RG must be modified.","tokens_in":47948,"feed_emoji":"🎲","tokens_out":7916,"duration_ms":59165,"temperature":0.7,"pith_summary":"The paper establishes that Wilson's exact renormalization group of 'incomplete integration' is, for the probability distribution of a field, a linear Fokker-Planck flow: every Fourier mode evolves as an independent Ornstein-Uhlenbeck process. Building on the Carosso interpretation, it reads the associated Langevin equation—the stochastic heat equation, also known as Edwards-Wilkinson dynamics—as a genuine infinitesimal coarse-graining of the field, with the noise representing the information discarded at each RG step. On a finite volume with periodic boundary conditions, the paper freezes the empirical magnetization (the zero-momentum Fourier coefficient) and adapts the Carosso RG to the remaining non-zero modes. It derives explicit RG flows for the intensive energy and for the generating function of the Fourier coefficients, so that the probability distribution of the empirical magnetization is conserved by construction and is obtained by gradually integrating out all non-zero momentum modes. If correct, this provides a stochastic-process representation of exact RG in which large-deviation properties of the magnetization emerge from a diffusion-like flow instead of a single high-dimensional integral.","feed_headline":"RG flows as stochastic coarsening yield magnetization large deviations","feed_subtitle":"A finite-volume Carosso stochastic RG gradually integrates out non-zero modes while freezing the empirical magnetization.","key_machinery":"The central object is the Ornstein-Uhlenbeck process for each Fourier coefficient of the field, with drift coefficient $-\\omega_L(k)\\hat\\phi$ and diffusion coefficient $D/(2L^d)$. The Laplacian eigenvalues $\\omega_L(k)=4\\pi^2 k^2/L^2$ provide the momentum-dependent damping; the zero-momentum mode is frozen by imposing the global constraint $\\int \\eta_t(x)\\,dx=0$ on the noise, which removes the zero-momentum noise component. This turns the exact RG flow into a diffusion process in the space of non-zero Fourier modes: the Fokker-Planck generator is separable, the finite-time propagator is a product of Gaussian kernels, and averaging over the noise realizes the gradual integration over the eliminated modes. The same machinery yields explicit flows for the intensive energy (through the transformation $P_t=\\exp(-L^d E_t)$) and for the generating function of the Fourier coefficients, making the conservation of $p_L(m_e)$ an identity at every RG time.","core_discovery":"The central claim is that the exact renormalization group of Wilson's incomplete integration is equivalent to a linear Fokker-Planck evolution for the renormalized probability distribution, in which the Fourier coefficients of the field undergo independent, non-identical Ornstein-Uhlenbeck processes. The corresponding Langevin stochastic differential equations—for the Carosso scheme, exactly the stochastic heat equation (Edwards-Wilkinson dynamics)—are not a mere change of variables in the partition function but genuine infinitesimal coarsening transformations, the continuous analog of block spins; the noise carries the information lost when modes are integrated out. On a finite volume $L^d$ with periodic boundary conditions, the paper imposes the constraint that the empirical magnetization $m_e$ remains frozen, so the zero-momentum coefficient is untouched and the RG flow acts only on the non-zero Fourier modes. The paper derives the resulting flows explicitly: the intensive energy $E_t$ satisfies the non-linear flow of Eq. 157, with its leading large-volume form in Eq. 158, and the generating function $W_t$ satisfies the linear flow of Eq. 166 with the finite-time expression of Eq. 163. These flows interpolate between the initial model and a factorized Gaussian steady state in which all non-zero modes have been integrated out and the only surviving information is the distribution $p_L(m_e)$ of the empirical magnetization itself.","pith_inferences":["The paper postulates Gaussian white noise rather than deriving it from the model's interactions; for a small non-Gaussian model (e.g., $\\phi^4$ theory on a finite lattice), one could measure the noise statistics produced by integrating out a thin momentum shell and check consistency with the assumed delta-correlated white noise.","The same construction should extend to freezing several low-lying Fourier coefficients, giving joint large-deviation functions for a set of intensive observables rather than only the zero mode.","Because the flow is linear in the generating function, it may be possible to reinterpret exact RG as a diffusion process whose stationary measure is the model's Gibbs measure, connecting to sampling schemes and generative diffusion models—connections the paper notes but does not develop."],"forward_implications":["The probability distribution of the empirical magnetization $p_L(m_e)$ can be obtained by running the Carosso flow until all non-zero Fourier modes reach their Gaussian steady state; the marginal at $t=\\infty$ is exactly $p_L(m_e)$, so the large-deviation rate function $i(m_e)$ follows from the stationary part of the flow.","At leading order in $1/L^d$, the RG flow for the intensive energy becomes a first-order Hamilton-Jacobi equation, so the rate function is obtained by optimizing the initial energy over the non-zero modes at fixed $m_e$.","The generating function of the Fourier coefficients obeys a linear first-order flow, meaning the scaled cumulants of the non-zero modes can be propagated exactly in time once the initial generating function is known.","The Markov property of the Ornstein-Uhlenbeck process gives a Chapman-Kolmogorov composition law for successive RG steps, so the whole flow can be built from elementary infinitesimal coarsening transformations—the continuous counterpart of iterative block-spin decimation."],"supporting_citations":[{"why":"Supplies the 'incomplete integration' exact RG scheme that the paper recasts as a Fokker-Planck flow for probabilities.","marker":"[19]"},{"why":"Introduced the continuity-equation perspective of RG as a change of variables conserving the partition function.","marker":"[34]"},{"why":"Developed the Wegner-Morris continuity equation for functional RG flows that the paper contrasts with the stochastic view.","marker":"[35]"},{"why":"Carosso's interpretation of Langevin SDEs associated to Fokker-Planck RG flows as genuine coarsening transformations.","marker":"[39]"},{"why":"Carosso's PhD manuscript containing the detailed Brownian-motion analogy for the noise and the lattice implementation.","marker":"[41]"},{"why":"Provides the large-deviation formalism (rate functions and scaled cumulants) used to define p_L(m_e) and i(m_e).","marker":"[12]"},{"why":"Recent exact functional RG analysis of the magnetization large deviations that provides the comparison target for the stochastic RG approach.","marker":"[13]"},{"why":"Reference for the Edwards-Wilkinson dynamics, which the paper identifies with the Carosso real-space Langevin equation.","marker":"[43]"}],"fun_headline_variants":["Exact RG as stochastic coarsening: magnetization large deviations","Wilson RG meets Langevin: coarsening yields magnetization tails","RG flow as Ornstein-Uhlenbeck: finite-volume magnetization large deviations","Stochastic exact RG: freezing magnetization, coarsening modes","Carosso RG: Langevin coarsening gives magnetization large deviations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the Gaussian white noise in the Langevin equation faithfully represents the information lost when high-momentum modes are integrated out; the paper assumes this noise form rather than deriving it from the model's interactions.","fun_headline_variants_meta":{"raw":{"variants":["Exact RG as stochastic coarsening: magnetization large deviations","Wilson RG meets Langevin: coarsening yields magnetization tails","RG flow as Ornstein-Uhlenbeck: finite-volume magnetization large deviations","Stochastic exact RG: freezing magnetization, coarsening modes","Carosso RG: Langevin coarsening gives magnetization large deviations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000634,"raw_usage":{"total_tokens":3060,"prompt_tokens":1214,"completion_tokens":1846,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":830,"completion_tokens_details":{"reasoning_tokens":1755}},"tokens_in":830,"tokens_out":1846,"duration_ms":12628,"temperature":1.0,"reasoning_tokens":1755,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T21:13:23.003138+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the Carosso flow for a simple interacting theory such as $\\phi^4$ on a small finite-volume lattice and compare the exact marginal $p_L(m_e)$ with the marginal obtained by integrating the flow to $t=\\infty$ with the assumed white noise. If the two disagree, the noise statistics are not those generated by the eliminated modes and the stochastic RG must be modified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the 'incomplete integration' exact RG scheme that the paper recasts as a Fokker-Planck flow for probabilities."},{"cited_title":"Touchette, Phys","cited_arxiv_id":null,"evidence_quote":"Introduced the continuity-equation perspective of RG as a change of variables conserving the partition function."},{"cited_title":"Balog, A","cited_arxiv_id":null,"evidence_quote":"Developed the Wegner-Morris continuity equation for functional RG flows that the paper contrasts with the stochastic view."},{"cited_title":"Wilson and J","cited_arxiv_id":null,"evidence_quote":"Carosso's PhD manuscript containing the detailed Brownian-motion analogy for the noise and the lattice implementation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Reference for the Edwards-Wilkinson dynamics, which the paper identifies with the Carosso real-space Langevin equation."}],"review_version":1}