{"id":"c8d26b50-b815-495e-ba07-40c729c959a8","arxiv_id":"1908.01641","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A least action principle with average-preserving variations yields an Euler-Lagrange condition with a deterministic component and links critical laws to McKean-Vlasov forward-backward SDEs.","lead":"This paper derives an Euler-Lagrange equation for laws of stochastic processes under perturbations that preserve the average of the process, yielding a deterministic term alongside a martingale. It connects critical laws for classical quadratic actions to McKean-Vlasov forward-backward stochastic differential equations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.1's proof outsources S-differentiability and the variational formula to Theorem 5.1 of the in-press, self-cited companion [31]; the central equivalence is only as secure as that unproven framework.","rationale":"The reader's weakest_assumption identifies the same point: Theorem 4.1 rests on Theorem 5.1 of [31] and the variational martingale characterization, neither of which is proved here. I agree with that assessment. The rest of the paper is internally coherent, Proposition 3.2 is plausible, and the explicit one-dimensional example checks out; these are real but partial supports. The central equivalence is nevertheless conditional on the companion framework being correct and applicable. Since my concern does not move the verdict away from the reader's conditional recommendation, I keep the verdict unchanged. I would not escalate to rejection because the cited framework is plausibly correct and the paper's own example provides a nontrivial consistency check.","tokens_in":11419,"tokens_out":28128,"duration_ms":303012,"concrete_test":"Independently re-derive the key variational formula: fix h in V^{∞,0}_ν and compute d/dε S((I+εh)_⋆ν)|_{ε=0} directly from (4.11), using only the definitions in Section 2, and show it equals Eν[<ξ,h>_H] with ξ_t = ∂_v L_t - ∫_0^t ∂_x L_s ds, under exactly (4.12)-(4.13). If the computation cannot be completed without additional hypotheses, for instance on the transformation of v^ν and α^ν under (I+εh)_⋆, then Theorem 4.1 remains unproved; if it succeeds, the external-framework concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing concern is the unproven import of Theorem 5.1 of [31]. In the proof of Theorem 4.1, after defining ξ, the text says: \"Under those conditions, the S-differentiability of S follows from Theorem 5.1 of [31]. Moreover, the proof of the latter also yields δSν[h] = Eν[<ξ,h>_H] for all h in L^2_a(ν,H_{0,0}).\" This is the step that turns the formal Euler-Lagrange expression into an actual derivative of the action (4.11); without it, assertion (i) of Theorem 4.1 has no connection to (4.17). Lemma 4.1 then supplies the equivalence, but its martingale side also cites the variational characterization from [23]/Proposition 1.1 of [31]. Neither result is proved in this paper, and Section 2 only records definitions and density statements, again citing [31]. Since [31] is an in-press article with overlapping authorship, the reader cannot check the framework from this preprint. This is a genuine gap in self-containedness, not a disagreement with consensus; if Theorem 5.1 of [31] requires assumptions beyond (4.12)-(4.13), for instance regularity of the characteristic maps v^ν and α^ν under the mapping (I+εh)_⋆, then the proof of Theorem 4.1 does not go through as written. The one-dimensional example is worked out and appears correct, which gives some support, but it does not cover the general mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a least-action principle for functionals of laws of continuous semimartingales, where variations are constrained to preserve the average of the Cameron-Martin path (average-preserving variation processes). The main result, Theorem 4.1, characterizes critical points of such action functionals: under regularity and growth conditions on the Lagrangian, a law is critical under all average-preserving variations with fixed endpoints if and only if the Euler-Lagrange expression equals the sum of a cadlag martingale and a deterministic measurable process. Section 5 connects this result to forward-backward SDEs of McKean-Vlasov type and gives a one-dimensional example. The proof relies heavily on the companion paper [31], an in-press article, for the variational framework, differentiability of the action, and the martingale characterization.","tokens_in":11723,"tokens_out":9268,"duration_ms":95205,"significance":"If Theorem 4.1 is valid, it provides a genuinely averaged-preserving Euler-Lagrange condition with an explicit deterministic correction term, which is a nontrivial and potentially useful tool in stochastic control and semimartingale optimization. The one-dimensional example is convincing: solving the SDE gives Y_t = B_t + e^t and X_1 ~ N(e-1, 7/3), matching the stated Gaussian law. The paper also gives a clean statement connecting the variational condition to forward-backward systems. However, the central result is conditional on the framework of [31], which is cited rather than proved, and Proposition 5.1 is stated without proof. The contribution is therefore significant but not yet self-contained.","major_comments":[{"comment":"The proof outsources the two load-bearing steps to the in-press companion [31]: the assertion that S is L^2_a(ν,H_{0,0})-differentiable under conditions (4.12)-(4.13), and the formula δS_ν[h] = E_ν[<ξ,h>_H] with ξ defined as in the proof. Both are said to follow from Theorem 5.1 of [31], whose statement and hypotheses are not reproduced. Since the reader cannot check that [31]'s theorem applies to the set S under the stated assumptions, the equivalence (i)⇔(ii) of Theorem 4.1 is not independently verifiable from this manuscript. Please state Theorem 5.1 of [31] in full, verify that conditions (4.12)-(4.13) imply its hypotheses, and give a proof or a published reference for the differentiability formula.","section":"Theorem 4.1, proof (Section 4)"},{"comment":"Lemma 4.1 also relies on a cited martingale characterization: the proof invokes the variational characterization of martingales 'see [23] or Proposition 1.1 of [31]' and the orthogonality of ∫_0^. (N_t - E_ν[N_0])dt to L^2_a(ν,H_{0,0}). This is a standard result, but the precise version needed for the representation ξ = ∫_0^. A_s ds + ∫_0^. N_s ds, with deterministic A, should be stated explicitly and proved in the present notation, since the pathwise integral of a martingale is not an element of H_{0,0} in general.","section":"Lemma 4.1 (Section 4)"},{"comment":"Proposition 5.1 is asserted without proof. This proposition is the concrete bridge between the variational theory and forward-backward systems, which is advertised as a central output of the paper. The equivalence between the existence of a solution to (5.19)-(5.20) and the variational condition (ii) should be proved directly, or the reduction to Theorem 4.1 should be carried out explicitly, including verification of the regularity conditions (4.12)-(4.13) and the endpoint constraints.","section":"Proposition 5.1 (Section 5)"}],"minor_comments":[{"comment":"The process ξ defined in the proof as ξ_t = ∫_0^t dot ξ_s ds does not necessarily have ξ_1 = 0, so it need not lie in H_{0,0}. To match the definition of L^2_a(ν,H_{0,0})-differentiability, one should replace ξ by its H_{0,0} projection; the difference is a linear deterministic term that can be absorbed into A^ν, but this step should be stated explicitly.","section":"Theorem 4.1, proof (Section 4)"},{"comment":"Equation (0.2) writes ∫_0^t ∂_q L(W_s, v^ν_s, α^ν_s), while the theorem uses ∂_x L; the notation in (0.2) appears to be a typo and should be aligned with the rest of the paper.","section":"Introduction, equation (0.2)"},{"comment":"The integrability condition '∫_0^1 |(σ.σ_t)^{i,j}_s(X)|ds < ∞' is difficult to parse; the notation (σ.σ_t) presumably denotes the matrix σ_s σ_s^T evaluated along X, but this should be written unambiguously.","section":"Section 5, Proposition 5.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is essentially an application of the framework developed in [31], an in-press paper with overlapping authorship. The main theorem depends on Theorem 5.1 of [31] for differentiability of the action and for the variational formula. This is a genuine self-containedness issue, not a criticism of the mathematics: the referee cannot verify the central claim from the submitted text. If the editorial policy permits dependence on an in-press companion, the paper may be publishable after the authors supply a full statement of the imported theorem and prove Proposition 5.1; otherwise the manuscript needs to be substantially expanded."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe one thing to know: this paper gives a genuinely new first-order condition for optimizing over semimartingale laws with a fixed mean, and the one-dimensional example checks out. The second thing: the main theorem is not self-contained. It outsources the differentiability of the action and the variational formula to Theorem 5.1 of the in-press companion [31], by the same author group. You cannot verify the central equivalence from this preprint alone.\n\nWhat's new: the average-preserving variation class A∞,0_ν, the density result in Proposition 3.2, and the Euler-Lagrange condition with a deterministic process A alongside the martingale term. The link to McKean-Vlasov FBSDEs in Section 5 is also new. Lemma 4.1 is clean and does the work of turning the martingale characterization into the equivalence, once differentiability is granted. Example 5.1 is a real check: the drift Y is explicitly a Brownian motion plus an exponential, and the resulting law is Gaussian with variance 7/3. That is concrete evidence the machinery is not empty.\n\nThe soft spots, in proportion. The load-bearing one is the outsourced proof. After defining ξ, the proof says 'the S-differentiability of S follows from Theorem 5.1 of [31]' and 'the proof of the latter also yields' the variational formula. That is exactly the step connecting the Euler-Lagrange expression to the derivative of the action. If Theorem 5.1 requires assumptions beyond (4.12)-(4.13), for example regularity of characteristic maps under (I+εh)_⋆, the proof of Theorem 4.1 fails. This is the core of the paper, not a minor gap. The example does not cover that generality, though it supports the final condition. A smaller point: Proposition 5.1 is stated more than proved; the FBSDE equivalence is asserted rather than demonstrated.\n\nThe paper is coherent and honest about its debts; the citation pattern is normal for a companion-paper setup. If [31] is solid, the main result is a useful contribution to stochastic calculus of variations. I would send it to peer review—it deserves referee time—but the verdict should be conditional on the companion framework being available and correct.\n\nFor a reader: someone working in stochastic control, Schrödinger problems, or semimartingale optimal transport. I wouldn't cite it in my own work until the dependency is resolved.\n\nRecommendation: accept for peer review, with a clear request that [31] be made available or the needed parts reproduced in an appendix.","headline":"New average-preserving least action principle with a real example, but the main theorem outsources its proof to an in-press companion; worth refereeing, but only after the dependency is resolved.","tokens_in":12243,"tokens_out":3199,"would_cite":false,"duration_ms":30750,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60H30","93E20"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that critical laws of semimartingale actions under average-preserving variations satisfy an Euler-Lagrange equation whose left side is a deterministic process plus a cadlag martingale.","keywords":["least action principle","average-preserving variations","semimartingale laws","Euler-Lagrange equations","intrinsic calculus of variations","forward-backward SDEs","mean-field SDEs","stochastic control"],"falsifier":"Take a smooth Lagrangian and a candidate extremal law $\\nu$ satisfying the endpoint and mean constraints, compute $\\partial_v L_t-\\int_0^t \\partial_x L_s ds$ along $\\nu$-typical paths, and test whether it can be written as a cadlag martingale plus a deterministic process; one example where this decomposition fails while $\\delta S_\\nu$ vanishes on all zero-mean variations would refute Theorem 4.1. A more basic check is to verify the density of the variation space in $L^2_a(\\nu,H_{0,0})$ for a non-Markovian law, since the proof imports that density from the companion paper.","tokens_in":11207,"feed_emoji":"⚖️","tokens_out":15084,"duration_ms":139943,"temperature":0.7,"pith_summary":"This paper establishes a least action principle for functionals of laws of continuous semimartingales, using perturbations that preserve the average of the path and fix both endpoints. It proves that a law is critical for such a variational problem exactly when the Euler-Lagrange expression, formed from the Lagrangian's derivatives with respect to drift and position, equals the sum of a deterministic measurable process and a cadlag martingale. This matters because it turns an infinite-dimensional optimization problem over path laws into a tractable dynamical condition. For the classical quadratic action it connects critical laws to forward-backward systems of mean-field SDEs, and it extends the least action viewpoint to laws that need not be absolutely continuous with respect to a fixed Gaussian reference measure.","feed_headline":"Stochastic least action splits into martingale and deterministic drift","feed_subtitle":"For classical quadratic costs it becomes a forward-backward mean-field system, linking variation to SDEs.","key_machinery":"The engine is the interplay of three objects. First, the space of average-preserving variation processes $$\\mathcal{A}^{\\infty,0}_\\nu=\\{h\\in\\mathcal{V}^{\\infty,0}_\\nu:\\mathbb{E}_\\nu[h]=0_H\\},$$ whose closure in $L^2_a(\\nu,H_{0,0})$ is exactly the zero-mean paths in $H_{0,0}$. Second, the intrinsic differential $\\delta S_\\nu$, defined by differentiating $S$ along the maps $\\mathbb{I}_W+\\epsilon h$, which makes calculus on laws possible without quasi-invariance. Third, the variational martingale characterization (Lemma 4.1): if $\\delta S_\\nu$ vanishes on the dense zero-mean variation space, the representing derivative $\\xi$ has the form $\\xi=\\int_0^\\cdot A^\\nu_s ds+\\int_0^\\cdot N^\\nu_s ds$, with $N^\\nu$ a cadlag martingale and $A^\\nu$ deterministic. Applying this to the action gradient $\\dot{\\xi}_t=\\partial_v L_t(W_t,v^\\nu_t,\\alpha^\\nu_t)-\\int_0^t \\partial_x L_s(W_s,v^\\nu_s,\\alpha^\\nu_s)ds$ delivers Theorem 4.1.","core_discovery":"The central claim is Theorem 4.1: for a regular Lagrangian $L$ satisfying the stated growth bounds, the action $S(\\nu)=\\mathbb{E}_\\nu[\\int_0^1 L_t(W_t,v^\\nu_t,\\alpha^\\nu_t)dt]$ is intrinsically differentiable at $\\nu$, and the first variation $\\delta S_\\nu[h]$ vanishes for every fixed-endpoint, zero-mean variation $h$ if and only if there are a cadlag martingale $(N^\\nu_t)$ and a deterministic measurable process $(A^\\nu_t)$ with $$\\partial_v L_t(W_t,v^\\nu_t,\\$\\alpha$^\\nu_t)-\\int_0^t \\partial_x L_s(W_s,v^\\nu_s,\\$\\alpha$^\\nu_s)ds = A^\\nu_t+N^\\nu_t,\\quad \\$\\lambda$\\text{-a.e.},\\ \\nu\\text{-a.s.}$$ The deterministic process is the signature of the average-preserving constraint: because variations with zero mean cannot shift the overall drift by a constant, the dual variable acquires a time-dependent deterministic component alongside the martingale.","pith_inferences":["The author does not say this, but the deterministic process $A^\\nu$ behaves like a Lagrange multiplier for the mean constraint, so the same martingale-plus-deterministic split should appear in any variational problem over path laws constrained only by mean and endpoints.","Applying the theorem to relative-entropy actions would likely reproduce entropic bridges; such a test would show whether the average-preserving least action gives a new constructive route to the classical bridge problem.","A possible testable extension is to check on simple nonlinear filters whether the extremality condition forces the innovations to be a martingale plus a deterministic process; if it does, the intrinsic calculus offers a variational route into the innovations problem of filtering.","In multi-dimensional systems, $A^\\nu_t$ depends only on time, which suggests it can be estimated from sample paths by subtracting the martingale part from the Euler-Lagrange left side and looking for a purely time-dependent remainder."],"forward_implications":["The Euler-Lagrange equation of Theorem 4.1 gives a concrete necessary and sufficient condition for a semimartingale law to be critical under average-preserving variations.","For the classical action $S(\\nu)=\\mathbb{E}_\\nu[\\int_0^1(|v^\\nu_t|^2/2+V(W_t))dt]$, criticality is equivalent to the existence of a solution of a forward-backward system of mean-field SDEs in which the backward component appears only through its centered version.","For suitable choices of the action, the extremal laws include laws of continuous semimartingales whose drift characteristic is an integrable process with independent increments.","The average-preserving constraint removes the deterministic drift as an admissible variation, which is why the Euler-Lagrange condition contains the extra deterministic process $A^\\nu$ and not only the martingale term.","The one-dimensional example shows that the theorem can be used to build explicit critical laws by solving a stochastic differential equation and reading off the deterministic and martingale components."],"supporting_citations":[{"why":"Supplies the intrinsic calculus of variations framework, including the density of the variation space in $L^2_a(\\nu,H_{0,0})$ and the differentiability theorem used in the proof of Theorem 4.1.","marker":"[31]"},{"why":"Supplies the variational characterization of martingales used to split the derivative into a deterministic process and a martingale.","marker":"[23]"},{"why":"Provides the classical functional-analytic arguments used to convert the variation condition into the differential Euler-Lagrange form.","marker":"[10]"},{"why":"Supplies the forward-backward SDE framework in which the critical points of classical actions are reinterpreted in Section 5.","marker":"[38]"}],"fun_headline_variants":["Average-preserving variation yields martingale plus deterministic drift","Least action for stochastic processes splits into martingale and drift","Variation with zero mean gives deterministic drift in extremal processes","From least action to McKean-Vlasov forward-backward SDEs","Stochastic least action: deterministic drift from averaging constraint"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole proof depends on the companion framework's claims that the allowed variations are dense enough and that the action is differentiable at the candidate law, together with the growth bounds on the Lagrangian; if any of that fails, the Euler-Lagrange equivalence collapses.","fun_headline_variants_meta":{"raw":{"variants":["Average-preserving variation yields martingale plus deterministic drift","Least action for stochastic processes splits into martingale and drift","Variation with zero mean gives deterministic drift in extremal processes","From least action to McKean-Vlasov forward-backward SDEs","Stochastic least action: deterministic drift from averaging constraint"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000628,"raw_usage":{"total_tokens":2860,"prompt_tokens":857,"completion_tokens":2003,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":473,"completion_tokens_details":{"reasoning_tokens":1918}},"tokens_in":473,"tokens_out":2003,"duration_ms":15388,"temperature":1.0,"reasoning_tokens":1918,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:07:24.554457+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a smooth Lagrangian and a candidate extremal law $\\nu$ satisfying the endpoint and mean constraints, compute $\\partial_v L_t-\\int_0^t \\partial_x L_s ds$ along $\\nu$-typical paths, and test whether it can be written as a cadlag martingale plus a deterministic process; one example where this decomposition fails while $\\delta S_\\nu$ vanishes on all zero-mean variations would refute Theorem 4.1. A more basic check is to verify the density of the variation space in $L^2_a(\\nu,H_{0,0})$ for a non-Markovian law, since the proof imports that density from the companion paper.","supporting_citations":[{"cited_title":", Cruzeiro, A.B","cited_arxiv_id":null,"evidence_quote":"Supplies the intrinsic calculus of variations framework, including the density of the variation space in $L^2_a(\\nu,H_{0,0})$ and the differentiability theorem used in the proof of Theorem 4.1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the variational characterization of martingales used to split the derivative into a deterministic process and a martingale."},{"cited_title":"Analyse fonctionnelle , Dunod, Paris 1999","cited_arxiv_id":null,"evidence_quote":"Provides the classical functional-analytic arguments used to convert the variation condition into the differential Euler-Lagrange form."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the forward-backward SDE framework in which the critical points of classical actions are reinterpreted in Section 5."}],"review_version":1}