{"id":"f298b68f-f580-45c1-9e65-533969820916","arxiv_id":"2412.20310","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For control-constrained semilinear parabolic tracking problems without Tikhonov regularization, the value function is Frechet differentiable in the initial value along the optimal trajectory, with gradient equal to the adjoint state.","lead":"This paper proves that, along an optimal trajectory of a semilinear parabolic optimal control problem with box constraints on the control, the value function is differentiable with respect to the initial state, and it identifies the derivative as the adjoint state. It does so under a quadratic growth assumption instead of the usual Tikhonov regularization, which may make the result useful for feedback design and policy iteration.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proof of Theorem 2.4's upper bound relies on an unproved adjoint-stability estimate ||p_eta(0)-p_bar(0)|| <= C||eta||, leaving the central Fréchet differentiability claim incomplete.","rationale":"The reader's stated weakest assumption is Assumption 2.2, which is indeed a strong, unverified hypothesis. However, my stress-test identifies a more specific and more immediately load-bearing gap: even granting Assumption 2.2 and the resulting state-stability Theorem 3.9, the proof of Theorem 2.4 needs the adjoint stability ||p_eta(0)-p_bar(0)|| <= C||eta|| to obtain the limsup inequality. This estimate is asserted as a 'fact' without proof, and it is not a formal consequence of the results stated earlier in the paper. The obstacle is that the adjoint difference's source contains a product of a state difference with p_bar; controlling it in L^2 requires p_bar in L^infty, which is not guaranteed by the stated embedding W(0,T) -> L^{2(d+2)/d}(Q), though it may follow from standard parabolic regularity if one adds a suitable hypothesis on the source. The same gap appears in Theorem 2.5, where p_h(tau+h) -> p_bar(tau) is used without proof. This does not necessarily invalidate the paper's results, but it means the central claim is not fully established as written. The reader's verdict of CONDITIONAL is therefore appropriate, and my read does not move it; the paper should add a lemma proving the adjoint-stability estimate, or state it as an explicit additional assumption, and should supply the deferred nonlinear-case arguments in Theorems 2.5 and 2.6.","tokens_in":21835,"tokens_out":14422,"duration_ms":130600,"concrete_test":"Prove the missing adjoint-stability estimate: show that for every minimizer u_eta of (P_{0,y0+eta}), the associated adjoint p_eta satisfies ||p_eta(0)-p_bar(0)||_H <= C||eta||_H for all sufficiently small eta. The proof must bound (f'(y_eta)-f'(y_bar)) p_bar in L^2(Q); this requires either a derivation of p_bar in L^infty(Q) from Assumption 2.1(iii) by parabolic maximal regularity, or an explicit extra hypothesis. If the estimate cannot be proved, Theorem 2.4's upper bound is unsupported. Independently, the same estimate should be supplied for p_h(tau+h) in Theorem 2.5.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In the proof of Theorem 2.4 (Section 4), the lower-bound half (liminf >= 0) is carried out in detail, but the opposite inequality is dismissed with the sentence: 'using integration by parts with the adjoint p_eta ... and using fact that ||p_eta(0)-p_bar(0)||_H <= c''||eta||_H.' This adjoint-stability estimate is never proved or cited. It is not an immediate consequence of Theorem 3.9: the difference P = p_eta - p_bar solves a backward parabolic equation with source (y_eta - y_bar) - (f'(y_eta) - f'(y_bar)) p_bar, and bounding the second term in L^2(Q) requires p_bar in L^infty(Q), which is not established anywhere (W(0,T) only embeds into L^{2(d+2)/d}(Q)). The same unproved convergence is used in Theorem 2.5, Step 3, where p_h(tau+h) -> p_bar(tau) is silently invoked. If this estimate fails, the limsup inequality can fail, so Fréchet differentiability, the paper's central claim, is not proven by the given arguments.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the value function of a control-constrained semilinear parabolic optimal control problem with a tracking-type objective and no Tikhonov regularization. It claims Fréchet differentiability of the value function with respect to the initial condition along the optimal trajectory (Theorem 2.4), one-sided and two-sided time differentiability under additional regularity of the optimal state (Theorem 2.5), joint differentiability in time and initial condition (Theorem 2.6), and discusses differentiability in a neighborhood of the trajectory. The proofs rely on a second-order growth assumption on the first and second variations (Assumption 2.2) and on stability estimates from previous work of the authors.","tokens_in":22087,"tokens_out":6974,"duration_ms":63369,"significance":"If the main theorems are correct, the paper would provide the first differentiability results for the value function of unregularized PDE-constrained optimal control problems under assumptions weaker than Tikhonov regularization, and the results are also new for ODE-constrained problems. The proof strategy based on joint growth of the first and second variations, together with the linearized-state stability estimate in Theorem 3.9, is an interesting and potentially useful framework. However, the current manuscript leaves a load-bearing adjoint-stability estimate unproved and provides only a sketch for the time-differentiability statements, so the central claims are not yet fully established.","major_comments":[{"comment":"The upper-bound half of the Fréchet differentiability proof is completed by the sentence: 'using integration by parts with the adjoint p_η ... and using fact that ||p_η(0)-p̄(0)||_H ≤ c''||η||_H for some constant c'' > 0 independent of η.' This adjoint-stability estimate is never proved or cited. It is not an immediate consequence of Theorem 3.9, because the difference P = p_η - p̄ solves a backward parabolic equation whose source contains the term (f'(y_η)-f'(ȳ))p̄, and bounding this term in L^2(Q) requires more regularity of p̄ than W(0,T) provides (W(0,T) embeds into L^{2(d+2)/d}(Q), not into L^∞(Q)). Since the limsup inequality depends on controlling the adjoint difference, this is a gap in the proof of the paper's main theorem. The authors should provide a full proof of this estimate or a precise reference.","section":"Section 4, proof of Theorem 2.4"},{"comment":"The passage to the limit h→0+ in the term ∫_Ω p_h(τ+h,x) [ȳ(τ+h)-ȳ(τ)]/h dx implicitly requires p_h(τ+h) → p̄(τ) in H (or an appropriate weak convergence with the stated normalization), but no adjoint-stability result for the family p_h is stated or cited. Moreover, the proof explicitly assumes f(y)=y and defers the case of general nonlinearities to 'standard arguments' without specifying the necessary estimates. Both points need to be addressed for Theorem 2.5 to be considered proven.","section":"Section 4, proof of Theorem 2.5, Step 3"},{"comment":"Theorem 2.6 is proved in a single sentence: 'The statement follows by the same arguments as in the proofs of Theorem 2.4 and Theorem 2.5.' Given that Theorem 2.5 is only proved for f(y)=y and already relies on the unproved adjoint convergence, the joint-differentiability claim needs at least a detailed sketch explaining how the state and time variations are coupled and which additional regularity is required. As written, the proof is insufficient to verify the theorem.","section":"Section 4, proof of Theorem 2.6"}],"minor_comments":[{"comment":"The interval 's ∈ [1, d+2/d)' is a typo for 's ∈ [1, (d+2)/d)' (and similarly 's < d+2/d' for 's < (d+2)/d'). The same notational ambiguity appears in the proof of Theorem 2.4.","section":"Section 3.2, Lemmas 3.7 and 3.8"},{"comment":"The displayed formula for A omits the term + ȳ^2 in the integrand. As written, ∫_Q [1/2(y_η^2 - ȳ^2) - ȳ y_η] dxdt is not necessarily nonnegative, whereas the correct quadratic remainder 1/2(y_η - ȳ)^2 is. The conclusion A ≥ 0 is correct once the typo is fixed.","section":"Section 4, equation after (4.2)"},{"comment":"The paper claims that Assumption 2.2 is 'generally applicable to a wide range of problems, including, for instance, certain tracking-type problems,' but no concrete example or explicit verification for any data is provided. A brief example or a reference where the condition is checked would help the reader assess the scope of the main results.","section":"Assumption 2.2"},{"comment":"The notation for the value function is inconsistent: Theorem 2.6 uses v while the rest of the paper uses υ. This should be unified.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a genuine open problem and the lower-bound half of Theorem 2.4 is carefully proved, but the upper-bound half depends on an unproved adjoint-stability estimate that is not obviously derivable from the provided material. Since this estimate is also used in Theorem 2.5, the central claims are currently incomplete. I recommend requiring a complete proof or reference for the adjoint-stability estimate, and a more detailed treatment of Theorems 2.5 and 2.6, before publication. If the estimate can be supplied, the manuscript would be a valuable contribution to the literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The genuinely new thing is the removal of the Tikhonov term: the paper proves Fréchet differentiability of the value function in the initial condition for control-constrained semilinear parabolic problems, with the gradient given by the adjoint state. That is a real step beyond [21,22], which need the regularization. The overall strategy is right. The lower-bound half of Theorem 2.4 is clean: the global growth lemma (3.4) plus integration by parts gives the liminf inequality directly, and Theorem 3.9, the state-stability estimate, is a solid ingredient.\n\nThe upper bound is where the argument goes soft. The proof says, in one sentence, that ||p_eta(0)-p_bar(0)||_H <= c''||eta||_H. That adjoint-stability estimate is neither proved nor cited. It is not an immediate consequence of the earlier theorems: the difference p_eta-p_bar solves a backward parabolic equation whose source contains (f'(y_eta)-f'(y_bar)) p_bar, and bounding that term in L^2 needs p_bar in L^infty(Q). The paper never establishes that; W(0,T) only embeds into L^{2(d+2)/d}(Q). So the limsup side of the main theorem is not proven by the text as written. The same unproved convergence appears again in Step 3 of Theorem 2.5.\n\nTwo more weaknesses, in proportion. Theorem 2.5 is explicitly proved only for f(y)=y, with the general nonlinearity deferred to 'standard' arguments. Theorem 2.6 is asserted in one sentence. And Assumption 2.2, the joint growth condition that drives everything, is imported from the authors' earlier work and never instantiated on a concrete example. It is a plausible sufficient condition, but the paper would be much stronger with a verified example or a precise citation of one.\n\nNone of this makes me think the result is false. The approach is coherent and the missing estimates could be proved with additional work. But as it stands, the central claim is not fully established. This paper deserves a serious referee, and I would send it to review, but the referee should require the adjoint-stability estimate to be proved or explicitly cited, and the sketches in 2.5/2.6 to be filled in. My bottom line: conditional acceptance after major revision.","headline":"Removing the Tikhonov term is genuinely new, but the main theorem's upper bound rests on an unproved adjoint-stability estimate, so the paper needs major revision before the claims are fully established.","tokens_in":22615,"tokens_out":3585,"would_cite":false,"duration_ms":31697,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49L99","35K58","49K40"],"pacs":[],"model":"deepseek-v4-flash","headline":"Under a second-order growth condition, the value function of a semilinear parabolic tracking problem is Frechet differentiable along the optimal trajectory, and its gradient is the adjoint state.","keywords":["value function","semilinear parabolic PDE","optimal control","control constraints","Fréchet differentiability","adjoint state","second-order sufficient conditions","solution stability"],"falsifier":"Take a concrete one-dimensional semilinear tracking problem (e.g., $f(y)=y$ with box constraints), verify Assumption 2.2 by direct computation, and check numerically whether the difference quotient $\\left( v(y_0+\\eta)-v(y_0)-\\langle \\bar p(0),\\eta \\rangle \\right)/\\|\\eta\\|_H$ tends to zero as $\\eta \\to 0$; a nonzero limit would contradict the theorem under its stated hypotheses, while a zero limit would confirm the mechanism.","tokens_in":21592,"feed_emoji":"🎛️","tokens_out":11539,"duration_ms":97450,"temperature":0.7,"pith_summary":"This paper establishes differentiability of the value function for an optimal control problem governed by a semilinear parabolic equation with box constraints on the control and no Tikhonov regularization. The main theorem says that if the reference control is a strict global minimizer satisfying a second-order growth condition on the cost (Assumption 2.2), then the value function is Frechet differentiable in the initial condition at every point of the optimal trajectory, and the gradient is the adjoint state. Under an additional time-differentiability assumption on the optimal state, the value function is also differentiable in time and jointly in both variables. The paper's significance is that differentiability of the value function is a prerequisite for HJB-based feedback synthesis and policy iteration, and this is the first such result for unregularized control-constrained parabolic problems under these weak assumptions.","feed_headline":"Adjoint state gives the value function gradient in parabolic control","feed_subtitle":"This differentiability is the missing step toward feedback control synthesis in unregularized problems.","key_machinery":"The engine is Assumption 2.2: for all small admissible directions $v$, $J'(\\bar u)v + \\tfrac{1}{2}J''(\\bar u)(v,v) \\ge c\\|z_v\\|_{L^2(Q)}^2$ with $c>0$, where $z_v$ solves the linearized state equation (2.3). This joint first-plus-second variation growth condition, imported from earlier stability theory, implies a global quadratic growth bound in state distance (Lemma 3.4), its restriction to shifted time intervals (Corollary 3.5), and the key stability estimate Theorem 3.9, which yields $\\|y^{\\tau+h,\\bar y(\\tau)+\\eta}_{u^{h,\\eta}} - \\bar y\\|_{L^2(Q_{\\tau+h})} \\le \\kappa (\\|\\bar y(\\tau+h)-\\bar y(\\tau)\\|_H + \\|\\eta\\|_H)$ for all small $h$ and $\\eta$. The adjoint state $\\bar p$ is the object that carries the gradient: the integration-by-parts identity for the linearized state, $\\int (\\bar y - y_Q) z = \\int \\bar p\\, \\delta u + \\langle \\bar p(\\tau), \\xi \\rangle$, is used twice (once with the reference data and once with the perturbed data) to squeeze the difference quotient of the value function between two expressions that converge to $\\langle \\bar p(\\tau), \\eta \\rangle$.","core_discovery":"The central claim is Theorem 2.4: under Assumptions 2.1 and 2.2, if the reference control $\\bar u$ is a strict global minimizer of the tracking problem (P), then for every $\\tau \\in [0,T]$ the value function $v(\\tau, \\cdot) : L^2(\\Omega) \\to \\mathbb{R}$ is Frechet differentiable at the reference optimal state $\\bar y(\\tau)$, and its gradient equals the adjoint state $\\bar p(\\tau)$. The proof proceeds by establishing the two-sided bound $\\limsup \\le 0 \\le \\liminf$ for the difference quotient in the initial condition. The upper bound uses the stability estimate Theorem 3.9 (which controls the $L^2$-distance between the perturbed optimal state and the reference state by a constant times the perturbation size), an $L^s$-$L^1$ estimate to control the error between the perturbed state and its linearization, and the adjoint identity to identify the leading term as the inner product with $\\bar p(0)$. The lower bound is obtained by reversing the roles of the reference and perturbed problems and using the adjoint of the perturbed problem. Theorems 2.5 and 2.6 extend the result to one-sided and two-sided time derivatives, and to joint differentiability in (time, state), under the additional hypothesis that the reference optimal state is time-differentiable in $L^2$ at the point.","pith_inferences":["The paper's proof suggests that the stability estimate Theorem 3.9 is the true workhorse: the same two-step template—state-distance stability plus adjoint-based squeezing—should transfer to other control-affine PDE problems (e.g., quasilinear or Navier-Stokes constraints) once an analogous stability estimate is available, even though the paper does not state this extension.","Because the paper verifies no concrete datum against Assumption 2.2, a natural next step is to check the assumption for a simple model (e.g., $f(y)=y$ with box constraints) and compute the value function numerically to confirm the predicted gradient, which would turn the abstract theorem into an implementable check.","The discussion of multiple minimizers suggests a plausible route to extend differentiability to problems with finitely many global minimizers, provided one can prove that the perturbed optimal states remain uniformly close to the set of reference optimal states; the paper leaves this as future work, so this is an inference.","If the value function is differentiable along the optimal trajectory, then the dynamic programming or HJB equation holds at those points, which could be used to construct locally stabilizing feedback laws; the paper mentions this motivation but does not derive such feedbacks."],"forward_implications":["For any $\\tau$, the value function $v(\\tau,\\cdot)$ is Frechet differentiable at the reference optimal state $\\bar y(\\tau)$, with gradient $\\bar p(\\tau)$.","If the reference optimal state is right (left) differentiable in time at $\\tau$, the value function has a one-sided derivative formula involving $\\bar p$ and the Lagrangian $L(\\bar y(\\tau))$; both one-sided derivatives coincide when the state is differentiable, giving full time differentiability.","At points where the optimal state is differentiable in time, the value function is differentiable jointly in (time, state) on $L^2(\\Omega)$.","The differentiability results hold for initial data in $H^1$ or $L^\\infty$ with the same arguments, and the $L^\\infty$ setting admits more general nonlinearities.","Under an additional control growth condition or a pointwise Legendre-Clebsch-type condition, the differentiability extends to a neighborhood of the optimal trajectory rather than only at the trajectory."],"supporting_citations":[{"why":"Introduces the joint first-plus-second variation growth condition that the paper imports as Assumption 2.2, and proves its sufficiency for solution stability.","marker":"[9]"},{"why":"Provides the parabolic solution-stability machinery, including the $L^s$-$L^1$ estimate used in Lemma 3.6 to bound the linearization error.","marker":"[14]"},{"why":"Establishes second-order sufficient conditions for parabolic control problems in terms of the linearized-state norm, motivating the form of Assumption 2.2.","marker":"[10]"},{"why":"Supplies the Taylor-type estimate used in Proposition 2.3 to derive the local quadratic growth bound from the growth assumption.","marker":"[11]"},{"why":"Proves continuous differentiability of the value function for Tikhonov-regularized parabolic problems; the present paper extends this to the unregularized setting.","marker":"[21]"},{"why":"Analogous value-function differentiability on $H^1$ initial data with regularization; the paper's discussion of $H^1$ and $L^\\infty$ variants builds on this comparison.","marker":"[22]"},{"why":"Standard reference for weak solutions of semilinear parabolic equations and the integration-by-parts formula used throughout the proofs.","marker":"[30]"}],"fun_headline_variants":["Value function gradient equals adjoint state","Adjoint state is value function gradient","Parabolic control: value function differentiability","Frechet differentiable value function in parabolic control","Parabolic optimal control: value function derivative"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is Assumption 2.2, a uniform lower bound $c>0$ on the first plus half the second variation of the cost in terms of the squared linearized-state norm, which the paper does not verify for any concrete datum but imports from earlier stability theory.","fun_headline_variants_meta":{"raw":{"variants":["Value function gradient equals adjoint state","Adjoint state is value function gradient","Parabolic control: value function differentiability","Frechet differentiable value function in parabolic control","Parabolic optimal control: value function derivative"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00077,"raw_usage":{"total_tokens":3412,"prompt_tokens":948,"completion_tokens":2464,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":564,"completion_tokens_details":{"reasoning_tokens":2400}},"tokens_in":564,"tokens_out":2464,"duration_ms":20191,"temperature":1.0,"reasoning_tokens":2400,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:24:03.611405+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a concrete one-dimensional semilinear tracking problem (e.g., $f(y)=y$ with box constraints), verify Assumption 2.2 by direct computation, and check numerically whether the difference quotient $\\left( v(y_0+\\eta)-v(y_0)-\\langle \\bar p(0),\\eta \\rangle \\right)/\\|\\eta\\|_H$ tends to zero as $\\eta \\to 0$; a nonzero limit would contradict the theorem under its stated hypotheses, while a zero limit would confirm the mechanism.","supporting_citations":[{"cited_title":"Cannarsa and H","cited_arxiv_id":null,"evidence_quote":"Introduces the joint first-plus-second variation growth condition that the paper imports as Assumption 2.2, and proves its sufficiency for solution stability."},{"cited_title":"Casas, A","cited_arxiv_id":null,"evidence_quote":"Establishes second-order sufficient conditions for parabolic control problems in terms of the linearized-state norm, motivating the form of Assumption 2.2."},{"cited_title":"Casas and M","cited_arxiv_id":null,"evidence_quote":"Supplies the Taylor-type estimate used in Proposition 2.3 to derive the local quadratic growth bound from the growth assumption."},{"cited_title":"Analysis of Unregularized Optimal Control Problems Constrained by the Boussinesq System","cited_arxiv_id":"2402.06873","evidence_quote":"Proves continuous differentiability of the value function for Tikhonov-regularized parabolic problems; the present paper extends this to the unregularized setting."},{"cited_title":"Kunisch and B","cited_arxiv_id":null,"evidence_quote":"Analogous value-function differentiability on $H^1$ initial data with regularization; the paper's discussion of $H^1$ and $L^\\infty$ variants builds on this comparison."},{"cited_title":"Preiss , Differentiability of L ipschitz functions on B anach spaces , J","cited_arxiv_id":null,"evidence_quote":"Standard reference for weak solutions of semilinear parabolic equations and the integration-by-parts formula used throughout the proofs."}],"review_version":1}