{"id":"c9917cf3-1abe-48b7-bb9e-a193c5c40358","arxiv_id":"2608.06923","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A new class of history-dependent differential stochastic variational inequalities is shown to be well-posed, sample-approximable, and transfer-stable.","lead":"This paper introduces a new class of stochastic systems where an ordinary differential equation is coupled to a history-dependent variational inequality, and proves that these systems have unique solutions, can be approximated from samples, and are stable under changes in the environment. The results provide a mathematical foundation for reusing precomputed responses across similar settings, demonstrated on a synthetic elderly-health monitoring benchmark.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 6's transfer-stability bound depends on M_Psi(t)<∞ and integrability of sqrt(M_Psi(t)), which are hidden in Lemma 5 rather than stated as hypotheses; for unbounded X or Xi the bound is vacuous, so the transfer guarantee is conditional on unstated global boundedness.","rationale":"The paper's well-posedness and SAA arguments are coherent, and I found no internal inconsistency in the monotonicity or Gronwall chains. The advertised quantitative transfer guarantee, Theorem 6, is the bridge to the application, and every constant in its bound flows through M_Psi(t). Because the theorem statements do not list compactness of X or boundedness of Ξ, nor the pointwise finiteness and integrability of sqrt(M_Psi(t)), the bound is incomplete rather than false: in natural unbounded data satisfying Assumption 1, both CY(t) and CT become infinite. The reader's weakest_assumption identified the hidden finiteness of M_Psi, and I agree; the additional requirement that sqrt(M_Psi) be integrable compounds the same concern. The health-monitoring experiments use compact and bounded data, so the fix is compatible with the applications and does not overturn the numerical conclusions. Thus the verdict remains conditional: the authors should state the missing hypotheses explicitly, or restrict the theorem to the compact/bounded setting that the applications actually use.","tokens_in":28344,"tokens_out":9416,"duration_ms":110130,"concrete_test":"Analytically instantiate the counterexample R=0, F(t,ξ,x,y)=y+ξ, Y(t,ξ)=[-1,1]^m, X=Ξ=R^n; verify Assumption 1(A1)-(A2) and the Hausdorff-Lipschitz condition hold, then compute M_Psi(t)=∞ and check that CY(t) in Lemma 5 and CT in Theorem 6 are infinite. Then amend Lemma 5 and Theorem 6 by adding explicit hypotheses (e.g., X compact and Ξ bounded, or M_Psi∈L^1 and sqrt(M_Psi)∈L^1) and re-run the Gronwall argument to confirm CT<∞; this settles whether the theorem requires an added boundedness/integrability hypothesis.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The transfer-stability theorem passes every quantitative estimate through Lemma 5, where M_Psi(t)=sup_{xi·∧t∈Λ, x∈X, y∈Y}||Psi(t,xi·∧t,x,y)|| is asserted to be finite inside the definition of CY(t). The proof of Theorem 6 then defines CT=exp(∫_0^T B(s)ds)·max{1,∫_0^T D(s)ds} with D(t)=ℓ_Phi(t)(1+C_Y(t)), and C_Y(t) contains sqrt(M_Psi(t)). Hence a finite CT requires not merely M_Psi(t)<∞ pointwise but also ∫_0^T sqrt(M_Psi(t))dt<∞. Neither condition is listed as a hypothesis in Section 4.2 or 4.3, and Assumption 1 does not imply it: A1 bounds R uniformly only for x in compact sets, while A2 is only Lipschitz in ξ, so with unbounded X or Ξ the supremum defining M_Psi is generally infinite. A concrete example is R=0, F(t,ξ,x,y)=y+ξ, Y(t,ξ)=[-1,1]^m, X=Ξ=R^n; Assumption A1-A2 and Hausdorff-Lipschitz Y hold, yet M_Psi(t)=∞ and both CY(t) and CT become infinite. The numerical experiments use X=[0,2] and bounded features, so the application is safe, but the theorem as stated is vacuous outside compact/bounded instances. This is the most load-bearing concern because every quantitative transfer guarantee degrades to ∞ if it lands.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a class of history-dependent differential stochastic variational inequalities (DSVIs) in which a projected ODE is coupled to a stochastic variational inequality whose data include the history of an exogenous process. Section 2 establishes existence, uniqueness, measurability, and state-Lipschitz continuity of the second-stage response, and hence well-posedness of the closed-loop trajectory under strong monotonicity, Lipschitz, and integrability conditions. Section 3 develops a sample average approximation of the expected drift and proves uniform almost-sure convergence of the approximate state trajectories. Section 4 contains the principal novelty: a local 1/2-Hölder estimate for parametric VIs with Hausdorff-Lipschitz moving feasible sets and a transfer-stability bound for first-stage trajectories in terms of the initial-state difference and the Wasserstein distance between exogenous path laws. Sections 5 and 6 provide numerical experiments on SAA convergence, source-to-target transfer stability, and an elderly-health monitoring application, where similarity-weighted reuse of source responses nearly matches the target-domain full-recomputation baseline while sharply reducing online latency.","tokens_in":28765,"tokens_out":10273,"duration_ms":108891,"significance":"If the theorems hold as stated, the paper is a useful contribution: it unifies several earlier DVI/SVI models, gives detailed existence and SAA proofs with explicit constants, and formulates a quantitative transfer-learning statement for a control-theoretic setting. The authors are careful to express all constants in Theorems 2, 4, and 6 through assumed quantities such as m_F, ℓ_F, ℓ_R, and W1 rather than fitted parameters, and the numerical study is reproducible via a linked repository. The paper also states limitations, including the single-user noise-robustness study and the exclusion of offline precomputation from the reported online runtime. The main caveat is that the transfer-stability theorem currently rests on an unstated global boundedness/integrability condition; once that condition is made explicit, the results would be a solid contribution to the DSVI and transfer-learning literature.","major_comments":[{"comment":"The transfer-stability result depends critically on the quantity MΨ(t), but its finiteness is asserted inside the definition of CY(t) in Lemma 5 rather than stated as a hypothesis, and Theorem 6 does not list any boundedness condition on X or Ξ that would imply it. Assumption 1 does not imply MΨ(t)<∞: for example, take X=R^n, Ξ=R^m, R≡0, F(t,ξ,x,y)=y+ξ, and Y(t,ξ)≡[-1,1]^m; then Assumption 1(A1)-(A2) and the Hausdorff-Lipschitz condition on Y hold, yet MΨ(t)=sup_{ξ∈R^m, y∈[-1,1]^m} ||y+ξ||=∞. Moreover, the proof of Theorem 6 needs not only pointwise finiteness of MΨ(t) but also integrability of sqrt(MΨ(t)) on [0,T], since D(t)=ℓΦ(t)(1+CY(t)) enters CT through ∫_0^T D(s)ds. Please add an explicit compactness or boundedness assumption on X and Ξ, or replace MΨ(t) by a local bound, and state the corresponding integrability condition as a hypothesis of Lemma 5 and Theorem 6.","section":"§4.2–4.3, Lemma 5 and Theorem 6"},{"comment":"The statement of Theorem 4 says 'Suppose that the assumptions of Theorem 2 hold', but its proof invokes Lemma 3, which additionally requires X to be compact. The paragraph before Lemma 3 states that compactness of X is assumed for the SAA convergence analysis, so this is likely intended as a standing assumption for Section 3; nevertheless, Theorem 4 as printed should include it explicitly. Otherwise the theorem statement is formally weaker than its hypotheses, and the SAA convergence claim is not justified without compactness.","section":"§3, Theorem 4"}],"minor_comments":[{"comment":"The notation qε=(ε,pε) is inconsistent with the displayed projection formula: the bound ||yε-y0||=sqrt(ε/(1+ε)) holds only if pε=√ε. Please define pε=√ε or write qε=(ε,√ε) explicitly.","section":"Example 1, §4.2"},{"comment":"The transfer experiments measure source-target similarity by feature-space distances d(v,u), whereas Theorem 6 is formulated in terms of W1 between full path laws. The connection between the theorem and the numerical transfer procedure is therefore heuristic; the text should acknowledge this gap rather than presenting the experiment as a direct instantiation of Theorem 6.","section":"§6.2.1"},{"comment":"The abstract says the transfer accuracy is 'close to the full-recomputation benchmark of 0.97', while Table 5 reports the full-recomputation baseline as 97.928% accuracy. The rounding to 0.97 is slightly loose; consider reporting 0.979 or writing 'over 0.97' to avoid an apparent inconsistency.","section":"Abstract and Table 5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the journal's scope in mathematical optimization and variational analysis. I do not see a circularity problem: the constants in Theorems 2, 4, and 6 are explicit in terms of the assumptions and are not obtained by fitting the numerical experiments. The overlap with the authors' earlier work is acknowledged through citations [10, 11, 17]. The main reason for major revision is the unstated boundedness/integrability condition on MΨ(t) in the transfer theorem; this is fixable by adding explicit hypotheses and should not require a full rewrite."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a real contribution, and the transfer-stability bound is worth publishing, but the main theorem is stated more broadly than the assumptions support. The paper's new content is genuine: a DSVI whose second-stage VI depends on the full history through the integral term involving R, reducing to the earlier models in [10, 11, 14] when R = 0; a well-posedness result built on strong monotonicity and Carathéodory ODE theory; an SAA convergence theorem; and a 1/2-Hölder stability estimate for parametric VIs with Hausdorff-Lipschitz moving feasible sets, leading to a trajectory bound in Wasserstein distance. The proofs are detailed and, as far as I checked, correct. The 1/2-Hölder exponent is not an artifact: the moving-set example in Section 4.2 shows why Lipschitz continuity fails, and the exponent is natural there.\n\nThe load-bearing soft spot is Lemma 5. The constant CY(t) is defined using MΨ(t) = sup ||Ψ||, and finiteness of MΨ(t) is asserted as part of the definition rather than listed as a hypothesis. Theorem 6 then needs ∫ sqrt(MΨ(t)) dt < ∞ for CT to be finite. Assumption 1 does not imply this: R is bounded only on compact x-sets and F is only Lipschitz in ξ, so with X or Ξ unbounded the supremum is typically infinite. A concrete counterexample: R = 0, F(t,ξ,x,y) = y + ξ, Y(t,ξ) = [−1,1]^m, X = Ξ = R^n satisfies A1–A2 and Hausdorff-Lipschitz Y, yet MΨ(t) = ∞ and both CY and CT blow up. The numerical experiments use compact X = [0,2] and bounded features, so the application is safe, but the theorem as stated is vacuous outside such cases. This is fixable: state explicitly that X and Ξ are compact (or that MΨ is finite with the needed integrability) and adjust the theorem statements accordingly.\n\nMinor issues: strong monotonicity with m_F > 0 drives all bounds; that is a real limitation but a standard one. The abstract's 'benchmark of 0.97' is sloppy against Table 5's 97.928%, and the health-monitoring experiments are synthetic with no error bars; the authors do note the noise-robustness study is one user. None of these change the theoretical contribution.\n\nWho this is for: researchers working on differential stochastic variational inequalities, stochastic generalized equations, or transfer learning in dynamic systems. The theory is solid enough to warrant a serious referee. I would send it out, with a request to fix the finiteness hypothesis and tighten the numerics.","headline":"Genuine extension of DSVI theory with a useful transfer-stability bound, but the main theorem's quantitative content is conditional on a finiteness assumption that is hidden in a definition rather than stated as a hypothesis.","tokens_in":29251,"tokens_out":2625,"would_cite":true,"duration_ms":25726,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C15","90C33","90C39"],"pacs":[],"model":"deepseek-v4-flash","headline":"Two similar stochastic environments produce close trajectories, the paper proves, giving a quantitative license for transfer learning in history-dependent differential stochastic variational inequalities.","keywords":["differential stochastic variational inequality","history-dependent response","transfer learning","sample average approximation","Wasserstein stability","parametric variational inequality","moving feasible set","elderly health monitoring"],"falsifier":"Fix identical initial states and drive the Wasserstein distance $W_1(\\mu^\\alpha,\\mu^\\beta)$ to zero while measuring $\\sup_{t \\in [0,T]}\\|x^\\alpha(t) - x^\\beta(t)\\|$: Theorem 6 requires this error to decay at least as fast as a constant times $W_1 + \\sqrt{W_1}$, so an experiment showing the error stays bounded away from zero as $W_1 \\to 0$, or decays at a strictly slower rate, would refute the bound. A crisp construction would embed the corner-tip feasible sets of the paper's Example 1 into the second stage and let the two path laws differ only in how often the path touches the corner.","tokens_in":28136,"feed_emoji":"⚖️","tokens_out":10515,"duration_ms":89881,"temperature":0.7,"pith_summary":"This paper proposes a class of history-dependent differential stochastic variational inequalities (DSVIs): an ordinary differential equation whose drift is the expected value of a feedback mapping involving a response defined by a variational inequality over the entire past path of an exogenous random process. It establishes that these systems are well posed — the stochastic response is unique, measurable, and Lipschitz in the current state, so the coupled trajectory exists and is unique — and that sample-average approximations of the drift converge uniformly almost surely. Its central result is a transfer-stability theorem: two such systems with different initial states and different laws for the exogenous process produce trajectories whose uniform distance is bounded by the initial-state difference plus the Wasserstein distance between the path laws and the square root of that distance. Because the bound is explicit and quantitative, it gives a theoretical justification for reusing responses computed in one stochastic environment as approximations in another similar one. The paper closes with an elderly-health monitoring case study in which similarity-weighted reuse of precomputed responses reaches accuracy close to the full-recomputation benchmark while cutting online batch runtime from 86 seconds to under one second.","feed_headline":"Proven: similar environments yield close trajectories","feed_subtitle":"New theorem licenses reusing responses across similar stochastic systems; health-monitoring batch time drops 139-fold.","key_machinery":"The load-bearing object is the history-dependent map $\\Psi(t,\\xi_{\\cdot\\wedge t},x,y) = \\int_0^t R(s,\\xi_s,x,y)\\,ds + F(t,\\xi_t,x,y)$, whose unique zero defines the second-stage response $y(t,\\xi_{\\cdot\\wedge t},x)$ as the solution of a parametric variational inequality over the moving feasible set $Y(t,\\xi_t)$. Under strong monotonicity of $F$ with constant $m_F > 0$ and Lipschitz problem data, the response is unique, measurable, and Lipschitz in the state; under a Hausdorff-Lipschitz assumption on the feasible-set map, the response changes by at most $C_Y(t)\\big(\\|\\xi^\\alpha - \\xi^\\beta\\|_\\infty + \\|\\xi^\\alpha - \\xi^\\beta\\|_\\infty^{1/2}\\big)$ between two histories. That square-root rate is carried through the expectation over path laws and integrated via Gronwall's inequality, which is how the theorem's $W_1 + W_1^{1/2}$ trajectory bound emerges.","core_discovery":"The paper's central discovery is Theorem 6: for two history-dependent DSVIs with identical data — same mappings $R$, $F$, $\\Phi$, same constraint sets, same horizon — but different initial states $x_0^\\alpha$, $x_0^\\beta$ and different exogenous path laws $\\mu^\\alpha$, $\\mu^\\beta$, the first-stage trajectories satisfy $$\\|x^\\$\\alpha$ - x^\\$\\beta$\\|_\\infty \\le C_T\\left(\\|x_0^\\$\\alpha$ - x_0^\\$\\beta$\\| + W_1(\\mu^\\$\\alpha$,\\mu^\\$\\beta$) + W_1(\\mu^\\$\\alpha$,\\mu^\\$\\beta$)^{1/2}\\right),$$ with a finite constant $C_T$ assembled from the regularity constants and the horizon. The exponent $1/2$ is not an artifact: a constructed example shows that feasible sets moving Lipschitz-continuously in Hausdorff distance can force the variational-inequality response to be only square-root-Hölder continuous in the path, and this rate propagates from the second-stage response through the drift to the trajectory. The paper reads this as a quantitative license for transfer learning: similar source and target environments and initial states yield close trajectories, so precomputed response trajectories from a source system can be reused for a target system with a controlled, explicit error.","pith_inferences":["The square-root term in Theorem 6 is plausibly sharp: because the paper's own Example 1 shows a Hausdorff-Lipschitz moving feasible set can force exactly $1/2$-Hölder response behavior, transfer error proportional to $\\sqrt{W_1}$ near nonsmooth feasible-set changes is likely unavoidable without extra structure.","A design principle follows that the paper leaves implicit: engineer the second-stage feasible set to vary smoothly, or precompute responses on a fixed grid of feasible sets, to convert the $1/2$-Hölder factor into a linear one and tighten the transfer bound.","Because the bound needs $\\sup\\|\\Psi\\|$ over all histories and states, it is vacuous for unbounded state spaces; replacing the supremum by an $L^p$ moment of $\\Psi$ under the path law would plausibly extend the transfer guarantee to unbounded systems.","The 139-fold latency reduction suggests the dominant online cost in such monitoring systems is second-stage response recomputation, so caching and similarity-weighted lookup are the effective levers for real-time deployment; this is worth confirming on non-synthetic cohorts."],"forward_implications":["Transfer learning across history-dependent DSVI systems carries a provable error certificate: when source and target differ by small initial-state and Wasserstein distances, the reused trajectory is close in uniform norm with the explicit constant $C_T$.","The $1/2$-Hölder rate tells practitioners to expect transfer error to shrink like the square root of the environment distance whenever feasible sets move with the environment, not linearly.","Sample-average approximation of the drift converges uniformly almost surely, so finite-sample computations faithfully approximate the exact expected-value dynamics as the sample size grows.","Response trajectories can be reused across time as well as users: refresh intervals of several minutes hold accuracy near the full-update benchmark while cutting recomputation cost by an order of magnitude.","In the elderly-health monitoring application, similarity-weighted reuse of precomputed responses reaches accuracy near the 97 percent full-recomputation benchmark while cutting online batch runtime from 86 seconds to about 0.62 seconds."],"supporting_citations":[{"why":"provides the existence and uniqueness theorems for strongly monotone variational inequalities that guarantee the second-stage response is well defined.","marker":"[19]"},{"why":"supplies the Carathéodory existence-and-uniqueness theorem and Gronwall lemma used to close every trajectory bound, including Theorem 6.","marker":"[38]"},{"why":"provides the empirical-process bracketing Glivenko–Cantelli theorem that drives the uniform almost-sure convergence of the SAA drift.","marker":"[41]"},{"why":"the quantitative stability framework for stochastic generalized equations that the transfer-stability analysis extends to history-dependent DSVIs.","marker":"[26]"},{"why":"establishes initial-condition dependence for deterministic differential variational inequalities, the precursor of the trajectory-stability result.","marker":"[34]"},{"why":"supplies the data-perturbation robustness analysis for stochastic generalized equations that motivates the transfer and noise experiments.","marker":"[21]"},{"why":"studies parameterized variational inequalities on moving sets under Lipschitz conditions and serves as the contrast showing why the weaker 1/2-Hölder regime is needed.","marker":"[17]"},{"why":"defines the Wasserstein metric used to compare exogenous path laws in the transfer bound.","marker":"[42]"},{"why":"introduces the parametric-optimization DSVI model and health-monitoring benchmark that this paper's history-dependent framework extends.","marker":"[11]"}],"fun_headline_variants":["Similar systems, close paths: transfer theorem","A stability bound for history-dependent DSVIs","Square-root Hölder transfer for stochastic systems","Reuse responses, cut batch runtime 139x","Quantitative license for transfer learning in DSVIs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on $F$ being strongly monotone in the response variable with a fixed positive constant $m_F$, which drives uniqueness of the response and every quantitative estimate, and on the response operator $\\Psi$ being uniformly bounded over all histories and states so that $M_\\Psi(t)$ is finite; if either premise fails, the stability and transfer bounds weaken or become vacuous.","fun_headline_variants_meta":{"raw":{"variants":["Similar systems, close paths: transfer theorem","A stability bound for history-dependent DSVIs","Square-root Hölder transfer for stochastic systems","Reuse responses, cut batch runtime 139x","Quantitative license for transfer learning in DSVIs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000244,"raw_usage":{"total_tokens":1600,"prompt_tokens":1084,"completion_tokens":516,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":700,"completion_tokens_details":{"reasoning_tokens":445}},"tokens_in":700,"tokens_out":516,"duration_ms":5904,"temperature":1.0,"reasoning_tokens":445,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:18:17.810717+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix identical initial states and drive the Wasserstein distance $W_1(\\mu^\\alpha,\\mu^\\beta)$ to zero while measuring $\\sup_{t \\in [0,T]}\\|x^\\alpha(t) - x^\\beta(t)\\|$: Theorem 6 requires this error to decay at least as fast as a constant times $W_1 + \\sqrt{W_1}$, so an experiment showing the error stays bounded away from zero as $W_1 \\to 0$, or decays at a strictly slower rate, would refute the bound. A crisp construction would embed the corner-tip feasible sets of the paper's Example 1 into the second stage and let the two path laws differ only in how often the path touches the corner.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the Carathéodory existence-and-uniqueness theorem and Gronwall lemma used to close every trajectory bound, including Theorem 6."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the empirical-process bracketing Glivenko–Cantelli theorem that drives the uniform almost-sure convergence of the SAA drift."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"the quantitative stability framework for stochastic generalized equations that the transfer-stability analysis extends to history-dependent DSVIs."},{"cited_title":"Guo and H","cited_arxiv_id":null,"evidence_quote":"supplies the data-perturbation robustness analysis for stochastic generalized equations that motivates the transfer and noise experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"studies parameterized variational inequalities on moving sets under Lipschitz conditions and serves as the contrast showing why the weaker 1/2-Hölder regime is needed."},{"cited_title":"Villani , Optimal Transport: Old and New , vol","cited_arxiv_id":null,"evidence_quote":"defines the Wasserstein metric used to compare exogenous path laws in the transfer bound."},{"cited_title":"Differential Stochastic Variational Inequalities with Parametric Optimization","cited_arxiv_id":"2508.15241","evidence_quote":"introduces the parametric-optimization DSVI model and health-monitoring benchmark that this paper's history-dependent framework extends."}],"review_version":1}