{"id":"2353c65c-8c6a-4e65-bd9e-ab7451e18c86","arxiv_id":"2607.13391","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A path-space Markov boundary holds exactly when the conditional law of two histories factorizes, and the failure gap is conditional mutual information that also bounds the thermodynamic cost of enforcing factorization.","lead":"This paper gives a unified mathematical framework for when, after conditioning on the full history of a boundary process, the histories of two other processes become independent. It connects this path-space condition to likelihood-factorisation, Girsanov control energies, and the thermodynamic free-energy cost of imposing separation.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central criterion rests on unproved domination of conditional path laws by clamped product references; this fails for noiseless/deterministic boundaries, so Theorem 5.3's scope is narrower than the paper's broad claims.","rationale":"The reader's weakest assumption identifies exactly the point on which the central factorization theorem rests: domination of the boundary-conditioned path law by a clamped product reference. I agree that this is the most load-bearing assumption. Theorem 5.3 is mathematically correct conditional on that assumption, but the paper neither proves it from the SDE data nor specifies regimes where it holds. The noiseless-boundary example shows the assumption can fail in a very simple, natural setting, so the central 'likelihood separation iff boundary' criterion is not as general as the title and abstract suggest. The reader's secondary findings are also real: Theorem 12.20's Girsanov norm is mismatched with Eq. (12.15), and Theorem 12.9 should require mutual absolute continuity; however these affect the thermodynamic maintenance bounds, not the core factorization theorem. Since the reader already returned CONDITIONAL for the same underlying reason, my read does not change the verdict. No change to the reader's assessment is needed.","tokens_in":26045,"tokens_out":7602,"duration_ms":81147,"concrete_test":"Test the domination hypothesis on a concrete Gaussian/Itô example where the boundary is a noiseless sum. Let Y and X be independent one-dimensional Brownian motions and set B_t = Y_t + X_t (no σ_B term). Take Q^b_Y and Q^b_X to be the clamped nondegenerate Wiener laws for Y and X. Fix b=0; then P^0 is supported on the affine subspace {Y = -X} in path space, while Q^0_Y ⊗ Q^0_X gives zero measure to that subspace, so P^0 is singular and no density H_0 exists. A finite-dimensional analogue is even simpler: Y,X ~ N(0,1) independent, B=Y+X; given B=0, the joint law is supported on the line y=-x, whereas the product of its nondegenerate marginals has full support and is singular. If this example satisfies the paper's hypotheses, Theorem 5.3 should apply; showing it does not demonstrates the missing domination condition is essential.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 5.3's iff criterion is explicitly conditional on the assumption that, for P_B-almost every boundary path b, the conditional law P^b is dominated by a clamped product reference Q^b_Y ⊗ Q^b_X (Section 6.1). The paper never derives this domination from the Itô coefficients, and it is not a harmless regularity condition. If the boundary is noiseless or degenerate—e.g., dB_t = (Y_t + X_t) dt with σ_B = 0—then conditioning on B=b forces Y[0,T] and X[0,T] onto a codimension-one set. Any clamped product law with nondegenerate diffusion assigns zero mass to that set, so P^b is singular with respect to Q^b_Y ⊗ Q^b_X. The Radon–Nikodym density H_b in Theorem 5.3 then does not exist, and the likelihood-separation criterion cannot even be formulated. This is not an internal inconsistency of the theorem as stated, but it is a load-bearing gap in the central claim: the abstract and summary present separation of a path log-likelihood as the general characterization of a path-space Markov boundary, yet the characterization applies only under a domination hypothesis that can fail in natural boundary models. The same issue affects Theorem 6.1 and the Bayes–Girsanov density (6.2), which all presuppose such domination without giving sufficient conditions in terms of the original SDE coefficients.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":26400,"tokens_out":12595,"duration_ms":120197,"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":[{"comment":"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.","section":"§12.6, Theorem 12.20, Eqs. (12.15)–(12.18)"},{"comment":"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.","section":"§12.2, Theorem 12.9"},{"comment":"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.","section":"§5, Theorem 5.3; §6.1, Theorem 6.1"}],"minor_comments":[{"comment":"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).","section":"§6.1"},{"comment":"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.","section":"§12.1, Theorem 12.11"},{"comment":"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.","section":"§12.6, Eq. (12.16)"},{"comment":"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.","section":"§12.6, Example 12.23"},{"comment":"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.","section":"§6.2, Proposition 6.2"},{"comment":"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.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper contains a substantial and mostly correct framework, but the two technical errors in Theorem 12.20 and Theorem 12.9 are load-bearing for the thermodynamic applications. They are local and correctable, so major revision is appropriate. The domination-assumption issue is acknowledged in the manuscript but the authors should add a discussion of its scope to avoid overclaiming. Overall, the paper has the potential to be a valuable contribution to the interface of stochastic analysis, information theory, and thermodynamics."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe 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.\n\nThe 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.\n\nThe 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.\n\nWho 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.\n\nRecommendation: 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.","headline":"Useful synthesis of path-space conditional independence; the new thermodynamic bounds have a concrete Girsanov metric error and the domination assumption is understated.","tokens_in":26859,"tokens_out":5877,"would_cite":false,"duration_ms":49501,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60H07","60J60","60J25","60G15","60G22","62F15","82C31"],"pacs":[],"model":"deepseek-v4-flash","headline":"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","keywords":["path space Markov boundary","conditional independence","Radon–Nikodym separability","Girsanov likelihood","Föllmer entropy identity","conditional mutual information","nonequilibrium steady state","sensor-actuator system"],"falsifier":"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.","tokens_in":1657,"feed_emoji":"🧩","tokens_out":1775,"duration_ms":68026,"temperature":0.7,"pith_summary":"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.","feed_headline":"Boundary path screens off system when likelihood factors into two","feed_subtitle":"One test—checking whether a log-likelihood separates—unifies boundary screening, control cost, and thermodynamic heat in one criterion.","key_machinery":"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","core_discovery":"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","pith_inferences":["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."],"forward_implications":["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."],"fun_headline_variants":["Two-path independence hinges on log-likelihood split","Markov boundary check: factorize the conditional density","When joint law factors, histories decouple","Factorization criterion for boundary screening","One test unifies boundary screening and control cost"],"cache_read_input_tokens":28032,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Two-path independence hinges on log-likelihood split","Markov boundary check: factorize the conditional density","When joint law factors, histories decouple","Factorization criterion for boundary screening","One test unifies boundary screening and control cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000129,"raw_usage":{"total_tokens":972,"prompt_tokens":769,"completion_tokens":203,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":135}},"tokens_in":513,"tokens_out":203,"duration_ms":2686,"temperature":1.0,"reasoning_tokens":135,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T05:19:31.111237+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[],"review_version":1}