{"id":"9bd14bc3-ecf1-4313-a6ab-e81bc7dd2505","arxiv_id":"2411.14031","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Using a truncated Carleman-weighted extremal problem and Uzawa-type augmented Lagrangian iterations, the paper computes numerical null controls for 2D/3D heat equations and 2D Stokes flow with reported terminal states near machine precision in 2D.","lead":"This paper introduces Lagrangian and Augmented Lagrangian algorithms, with a truncated Carleman weight, for numerically computing null controls of parabolic PDEs such as the heat and Stokes systems. The approach is demonstrated on 2D and 3D examples and includes convergence results for the truncation step.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported terminal norms are computed with the same discrete operators used to build the control; no error analysis or mesh-convergence study links them to the continuous parabolic null-controllability problem.","rationale":"The reader's weakest assumption is precisely the one that is most load-bearing for the paper's central claim. The paper proves a clean continuous convergence result for the truncated problem, and the algorithms are plausible applications of standard augmented-Lagrangian theory. But the numerical terminal states are produced by a discretized solution operator M and its adjoint M*, and the paper supplies no error analysis relating the discrete saddle-point solution to the continuous truncated problem. The text itself flags this only in passing ('standard finite dimensional reduction process'), and the reported experiments contain no spatial mesh-convergence study, no code/data release, and no independent verification that the computed control works on a finer discretization. This is not an internal contradiction or a disagreement with consensus; it is an evidential gap separating the discrete computations from the PDE claim. The recommended conditional verdict remains appropriate: the theoretical part appears sound, but the numerical validation must be made reproducible and the discrete-to-continuous link must be either proven or empirically demonstrated before the methodology can be accepted as solving the PDE null-controllability problem. I therefore keep the reader's CONDITIONAL verdict unchanged.","tokens_in":19979,"tokens_out":7335,"duration_ms":80082,"concrete_test":"Fix the Test #1 setting (R=10^5, K=0.1, ω=(0.2,0.5)×(0.2,0.6), Nt=80, 2358 triangles) and run ALG3 to convergence to obtain v_{h,Δt}. Then, keeping this control fixed, solve the state equation (1.1) on a sequence of uniformly refined spatial meshes and time steps (h,Δt → h/2,Δt/2; h/4,Δt/4) and with an independent solver; measure ||y(·,T)||_{L^2} and ||y(·,T)||_{H^1_0}. If these norms grow by orders of magnitude under refinement, or fail to converge to a limit compatible with the continuous bound C|log R|/R, the observed 1e-13 terminal norm is a discrete artifact and the claimed PDE-level null controllability is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central numerical claim is that ALG 1–4 solve the truncated extremal problem (2.5) and thereby produce null controls for the heat and Stokes parabolic PDEs. The theory in Section 2 gives a continuous convergence rate, but the objects actually computed in Section 6 are saddle points of a fully discrete problem: M and M* are replaced by finite-element plus implicit Euler/Gear approximations, as stated at the end of Section 4 and again in Section 5 ('achieved through a standard finite dimensional reduction process'). No theorem or error estimate connects the discrete saddle point to the continuous truncated problem. In particular, there is no bound of the form\n\n  ||y_R^{h,Δt}(·,T) - y_R(·,T)||_{H^1_0} ≤ ε(h,Δt,R), with ε → 0 as h,Δt → 0,\n\nnor any discrete analogue of Proposition 2.6. Table 3 varies Nt while leaving the spatial mesh essentially uncontrolled, and no test re-solves the state equation with the computed control on a finer or independent discretization. The reported terminal norms (10^{-13} in 2D, 10^{-7} in 3D, 10^{-9} for Stokes) therefore measure only the discrete dynamics. A finite-dimensional system can be driven to zero even when the continuous system cannot, or when the discretization is too coarse to represent the true control. Without a mesh-refinement study or a post-processing check, the headline claim 'numerical null controllability of parabolic PDEs' is not established; the reader's conditional verdict correctly identifies this as the weakest link.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a Lagrangian and Augmented Lagrangian approach to compute numerical null controls for linear parabolic equations and the Stokes system. The authors consider the Fursikov–Imanuvilov weighted extremal problem (1.4), introduce a truncated weight ρ_R = min(R, ρ) in Section 2.1, and prove that the minimizers v_R of the truncated problem converge strongly in U to the minimizer of (1.4) as R → ∞ (Proposition 2.5), with the terminal state satisfying an H^1_0 bound of order |log R|/R (Proposition 2.6). They then derive saddle-point reformulations in Sections 3–4 and present four iterative algorithms (Uzawa, conjugate-gradient Uzawa, and augmented Lagrangian variants), together with numerical experiments for 2D and 3D heat equations and a 2D Stokes system in Section 6. Reported terminal norms are on the order of 10^{-13} (2D heat), 10^{-7} (3D heat), and 10^{-9} (Stokes).","tokens_in":20324,"tokens_out":7307,"duration_ms":65310,"significance":"The paper addresses a real numerical bottleneck: constructing null controls for parabolic PDEs in high dimensions without relying on C1 finite elements or flatness-based expansions. The truncation idea is natural, the convergence proof of Proposition 2.5 is convincing, and Proposition 2.6 gives a quantitative rate that is plausible and consistent with the R-scaling in Table 1. The algorithms are clearly described and the experiments are reproducible in principle with FreeFem++. However, the central numerical validation is incomplete: the computed terminal norms are produced by the same discrete operators used to construct the controls, and no error analysis or mesh-convergence study connects the fully discrete solution to the continuous truncated problem. The significance of the method will be established only after this gap is addressed.","major_comments":[{"comment":"The fully discrete computations are not linked to the continuous truncated problem. At the end of Section 4 the authors state that the action of M and M* is approximated 'through a standard finite dimensional reduction process' (implicit Euler/Gear in time and FEM in space), but no error estimate is given, such as ||y_R^{h,Δt}(·,T) − y_R(·,T)||_{H^1_0} ≤ ε(h,Δt,R) with ε → 0 as h,Δt → 0. Consequently, the reported final-time norms in Test #1 (1.72×10^{-12}), Table 1, Table 3, Test #4 (2.10×10^{-7}), and Test #5 (5.50×10^{-9}) are computed with the same discrete operators that produced the control and certify only the discrete dynamics. Table 3 varies Nt but does not refine the spatial mesh in a controlled way, and no experiment re-solves the state equation on a finer or independent discretization. To support the claim of numerical null controllability of the PDE, the authors should add either a discrete error analysis or a systematic mesh/time refinement study with an a posteriori check that the computed control drives the continuous (or at least a finer-discrete) state to zero.","section":"End of Section 4; Section 6"},{"comment":"The proof of Proposition 2.6 has notation slips and a compressed energy estimate that need to be fixed. After defining z := ψ_R ρ_R y_R, the equation for z is written with χ_R instead of ψ_R, and the subsequent bound uses q_R(·,T) where z(·,T) is meant. The step from ∫|Lz|^2 ≤ C J(v̂)/(T2,R − T1,R)^2 to the parabolic regularity estimate ||z||_{C^0([0,T];H^1_0) ∩ L^2(H^1_0∩H^2)} ≤ C |log R|^2 J(v̂)^{1/2} is not demonstrated; the time scales of the cut-off and the boundedness of ρ_R near T require a careful statement. Since Proposition 2.6 provides the theoretical convergence rate that motivates the choice of R in the numerics, the proof should be completed and rewritten.","section":"Section 2.2, Proof of Proposition 2.6"},{"comment":"The theoretical convergence condition for Uzawa's algorithm in Section 4.1 requires the optimal steps r_k to satisfy r_k < 2C1(R)/C0^2 ~ R^{-2}, which would imply a severe dependence of the iteration count on R. However, Table 2 reports an essentially constant number of iterations (~3000) for R ranging from 10^3 to 10^8. The authors do not provide a convergence analysis for the augmented Lagrangian variant (ALG 3) actually used in the experiments, nor do they explain this apparent discrepancy. This is not a fatal flaw, but it should be addressed to support the advertised robustness of the algorithms.","section":"Section 4.1; Table 2"}],"minor_comments":[{"comment":"In Section 5, the augmented Lagrangian is defined as LR,K = LR + 1/2 ||M v + ȳ − y||^2, missing the factor K; this is inconsistent with Section 3.2 and with Algorithm 4, which uses K in the linear system. It should be K/2.","section":"Section 5"},{"comment":"The title of Algorithm 4 has a missing closing parenthesis, and the procedure is named Uzawa03 instead of Uzawa04.","section":"Algorithm 4"},{"comment":"In the convergence branch of Algorithms 1 and 3, the variable err is not updated and no break is executed, so the loop condition may not stop the iteration when the tolerance is first met; an explicit exit should be added.","section":"Algorithms 1 and 3"},{"comment":"In Table 1, the entry '1.319610−13' should read '1.3196 × 10^{-13}'.","section":"Table 1"},{"comment":"The sentence 'the uniform convergence of the control and the controlled state is clear' is not supported by Table 3, since the L2(Q)-norms of y and v still vary with Nt; please report a convergence sequence for Nt → ∞ or rephrase.","section":"Section 6.3"},{"comment":"There are typos: 'nonzeo' in Section 1.2 should be 'nonzero', and 'paremater' in Section 7 should be 'parameter'.","section":"Sections 1.2 and 7"}],"recommendation":"major_revision","confidential_remarks":"The main risk is that the numerical experiments, while suggestive, are not yet evidence for continuous null controllability unless the discrete-to-continuous gap is addressed. The paper's own text acknowledges the finite-dimensional reduction without analysis. I would recommend asking for a mesh-convergence study or a discrete error estimate before considering publication. The paper is otherwise within scope for math.OC."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new part is Section 2: truncating the unbounded Carleman weight via ρ_R = min(R, ρ). Proposition 2.5 (strong convergence of truncated minimizers) is proved by a clean compactness argument, and Proposition 2.6 (terminal bound O(|log R|/R)) is plausible, despite notation slips in the proof. That theoretical core is worth having, and the truncation idea is simple enough to be useful beyond this paper. The Lagrangian and Augmented Lagrangian machinery is classical, but applying it to this truncated formulation, with Uzawa and conjugate-gradient variants (ALG 1–4), is a reasonable contribution, especially for 3D where space-time FEM is impractical.\n\nThe soft spot is exactly what the stress-test says: the computed terminal norms (1e-13 in 2D, 1e-7 in 3D, 1e-9 in Stokes) are produced by the same discrete operators used to construct the control. There is no mesh-convergence study, no independent discretization check, no error bars, and no released code or data. So the numerics do not yet demonstrate null controllability of the continuous PDEs; they demonstrate that a finite-dimensional system can be driven to zero. That is a real gap, and the paper does not try to hide it—the finite-dimensional reduction is mentioned plainly—but it is never analyzed. Table 3 varies Nt while the spatial mesh is basically uncontrolled, and no experiment re-solves the state on a finer grid with the computed control.\n\nI would not call this fatal. The theory is sound on its own terms, and the missing piece is a discrete-to-continuous error estimate or at least a careful numerical verification. The citations are appropriate, with no circularity: the Carleman inequality is imported from Fursikov-Imanuvilov and Fernández-Cara–Guerrero, not from the authors' own prior work. The paper deserves a serious referee. It should go to review, with the referee asked to require a spatial mesh-refinement study, a post-processing check of the state equation on an independent discretization, and ideally a discrete convergence theorem. If the authors can provide that, the truncation result plus algorithms would make a solid contribution to numerical controllability.","headline":"New truncation trick with a clean convergence proof, but the numerics lack a discrete error analysis and reproducibility, so the central computational claim remains unverified.","tokens_in":20860,"tokens_out":3163,"would_cite":true,"duration_ms":33643,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35K05","49J20","65K10","65M60","93B05"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that truncating the time-blow-up weight turns numerical null controllability of parabolic PDEs into a well-posed saddle-point problem that Uzawa-style algorithms solve, with final-state error shrinking like log R over R.","keywords":["null controllability","heat equation","Stokes system","Lagrangian methods","Augmented Lagrangian","Uzawa algorithm","Carleman weights","finite element methods"],"falsifier":"Fix a moderate $R$ and solve the 2D heat test on a sequence of refined meshes and time steps. If the computed terminal norm plateaus at a mesh-dependent floor, or if increasing $R$ at fixed discretization stops shrinking the terminal norm, then the discrete saddle point is not tracking the continuous truncated problem.","tokens_in":19729,"feed_emoji":"","tokens_out":10288,"duration_ms":89161,"temperature":0.7,"pith_summary":"This paper tries to establish that null controllability for linear parabolic equations and systems can be solved numerically by applying classical Lagrangian and Augmented Lagrangian methods to a truncated weighted extremal problem. The key move is to replace the unbounded Carleman weight by $\\rho_R=\\min(R,\\rho)$, which makes the dual problem coercive and numerically well behaved while preserving the original problem in the limit. The authors prove that the truncated controls converge strongly to the minimizer of the original extremal problem as $R\\to\\infty$, and that the terminal state obeys $\\|y_R(\\cdot,T)\\|_{H^1_0(\\Omega)}\\le C\\,|\\log R|/R$. If the claim is right, standard finite-element and time-stepping tools are enough to compute accurate null controls for 2D and 3D heat equations and for the 2D Stokes system, with reported 2D terminal norms near $10^{-13}$.","feed_headline":"Lagrangian method computes heat and Stokes null controls","feed_subtitle":"Truncating the blow-up time weight turns zeroing the final state into a tractable saddle-point problem.","key_machinery":"The load-bearing objects are the Carleman weights $\\rho=e^{\\alpha}$ and $\\rho_0=(T-t)^{3/2}\\rho$, the truncated weight $\\rho_R=\\min(R,\\rho)$, and the solution operators $M$ and $M^*$: $Mv$ is the state driven from zero by control $v$, while $M^*q$ is the trace on $Q_\\omega$ of the backward adjoint state. These operators turn the constrained extremal problem into the unconstrained dual problem of minimizing $J_R^*(q)=\\frac12\\int_Q(\\rho_R^{-2}|q|^2+\\rho_0^{-2}|M^*q|^2+2q\\bar y)\\,dxdt$ over $q\\in L^2(Q)$, whose optimality conditions are $y_R=-\\rho_R^{-2}q_R$ and $v_R=\\rho_0^{-2}M^*q_R$. The algorithms are gradient, conjugate-gradient, and Augmented Lagrangian iterations on this dual functional; the Augmented Lagrangian versions solve an inner linear system with a Gauss-Seidel sweep, and the convergence-rate estimate uses a cut-off function supported near $t=T$ to convert the $|\\log R|$ growth of the weights into the $|\\log R|/R$ terminal bound.","core_discovery":"The central claim is that for each fixed $R>0$, the truncated extremal problem has a unique solution $(y_R,v_R)$, and Propositions 2.5 and 2.6 give $v_R\\to \\hat v$ strongly in $U$ and $\\|y_R(\\cdot,T)\\|_{H^1_0(\\Omega)}\\le C|\\log R|/R$, where $(\\hat y,\\hat v)$ is the minimizer of the original weighted extremal problem (1.4). The argument is carried by a saddle-point duality: the state and control are recovered from the dual variable by $y_R=-\\rho_R^{-2}q_R$ and $v_R=\\rho_0^{-2}M^*q_R$, so solving the truncated problem reduces to minimizing a strictly convex quadratic functional over $L^2(Q)$. The paper further claims that Uzawa, conjugate-gradient, and Augmented Lagrangian iterations (ALG 1-4) solve this dual problem efficiently, and the reported experiments show final $L^2$ norms of about $1.7\\times10^{-12}$ for a 2D heat equation, $2.1\\times10^{-7}$ for a 3D heat equation, and $5.5\\times10^{-9}$ for the 2D Stokes system.","pith_inferences":["A testable extension the paper does not pursue is a grid-convergence study at fixed $R$: if the terminal norm stops improving once mesh and time steps are refined, the $10^{-13}$ numbers would be discrete artifacts rather than evidence for the continuous estimate.","Because iteration counts are insensitive to $R$, the truncation parameter could be chosen adaptively per mesh level rather than fixed at values up to $10^8$, potentially reducing cost on coarser grids.","The same saddle-point machinery could be applied to boundary controls or semilinear problems, but the Carleman-weight estimates behind Propositions 2.5 and 2.6 would need to be re-proved since the cut-off argument relies on the specific form $\\rho_0=(T-t)^{3/2}\\rho$.","If the rate $|\\log R|/R$ is treated as a practical accuracy guide, then choosing $R$ from a target tolerance, roughly $R\\sim |\\log\\varepsilon|/\\varepsilon$, would give a self-tuned stopping rule for the truncation parameter."],"forward_implications":["ALG 3 recovers the predicted $|\\log R|/R$ scaling: Table 1 reports $\\|y(\\cdot,T)\\|_{L^2}$ dropping from about $1.3\\times10^{-9}$ at $R=10^3$ to $3.4\\times10^{-19}$ at $R=10^8$.","The number of Uzawa iterations stays essentially constant, near 3,000, as $R$ grows from $10^3$ to $10^8$, so taking a very large truncation parameter is not computationally expensive.","The same Augmented Lagrangian formulation handles 2D heat, 3D heat, and 2D Stokes, with final $L^2$ norms of order $10^{-12}$, $10^{-7}$, and $10^{-9}$, respectively.","Varying the number of time steps $N_t$ from 20 to 200 changes the final-state norm only mildly, staying around $10^{-13}$, so the computed controls are not highly sensitive to the time grid.","The authors state the method carries over to other parabolic systems and boundary conditions, and leave wave and Schr\\\"odinger controllability as open directions."],"supporting_citations":[{"why":"Originates the weighted extremal formulation (1.4) and supplies the Carleman weights and controllability theory on which the truncation is based.","marker":"[10]"},{"why":"Provides the convex-analysis duality and saddle-point results used to reformulate the truncated problem as the unconstrained dual problem (3.1).","marker":"[5]"},{"why":"Documents the ill-posedness of minimal-$L^2$ null control and motivates the regularized and truncated approach taken here.","marker":"[3]"},{"why":"Sets up the space-time variational strategy and strong convergence of approximations that the Lagrangian approach adapts.","marker":"[7]"},{"why":"Supplies global Carleman inequalities for parabolic systems that justify the observability estimates and the choice of weights.","marker":"[6]"},{"why":"Provides the conjugate-gradient comparison and the observation of lack of uniform coercivity that the reported iteration counts echo.","marker":"[8]"},{"why":"Gives the Augmented Lagrangian methodology used in ALG 3 and ALG 4.","marker":"[9]"},{"why":"Is the finite-element software used to run the reported numerical experiments.","marker":"[13]"}],"fun_headline_variants":["Saddle-point solver zeros heat and Stokes states","Lagrangian method computes exact null controls","Carleman-weighted trick zeros parabolic PDEs","Numerical null control via augmented Lagrangian","Heat and Stokes zeroed by Lagrangian saddle point"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the finite-element and time-stepping discretizations of the PDE solution operators are accurate enough that the discrete saddle point still represents the continuous truncated problem; the paper describes the discretization but supplies no error analysis linking the discrete and continuous solutions.","fun_headline_variants_meta":{"raw":{"variants":["Saddle-point solver zeros heat and Stokes states","Lagrangian method computes exact null controls","Carleman-weighted trick zeros parabolic PDEs","Numerical null control via augmented Lagrangian","Heat and Stokes zeroed by Lagrangian saddle point"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000164,"raw_usage":{"total_tokens":1231,"prompt_tokens":917,"completion_tokens":314,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":247}},"tokens_in":533,"tokens_out":314,"duration_ms":3258,"temperature":1.0,"reasoning_tokens":247,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:36:53.141915+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix a moderate $R$ and solve the 2D heat test on a sequence of refined meshes and time steps. If the computed terminal norm plateaus at a mesh-dependent floor, or if increasing $R$ at fixed discretization stops shrinking the terminal norm, then the discrete saddle point is not tracking the continuous truncated problem.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Originates the weighted extremal formulation (1.4) and supplies the Carleman weights and controllability theory on which the truncation is based."},{"cited_title":"Ekeland and R","cited_arxiv_id":null,"evidence_quote":"Provides the convex-analysis duality and saddle-point results used to reformulate the truncated problem as the unconstrained dual problem (3.1)."},{"cited_title":"Carthel, R","cited_arxiv_id":null,"evidence_quote":"Documents the ill-posedness of minimal-$L^2$ null control and motivates the regularized and truncated approach taken here."},{"cited_title":"Fern´ andez-Cara and A","cited_arxiv_id":null,"evidence_quote":"Sets up the space-time variational strategy and strong convergence of approximations that the Lagrangian approach adapts."},{"cited_title":"Fern´ andez-Cara and S","cited_arxiv_id":null,"evidence_quote":"Supplies global Carleman inequalities for parabolic systems that justify the observability estimates and the choice of weights."},{"cited_title":"Fern´ andez-Cara and A","cited_arxiv_id":null,"evidence_quote":"Provides the conjugate-gradient comparison and the observation of lack of uniform coercivity that the reported iteration counts echo."},{"cited_title":"Fortin and R","cited_arxiv_id":null,"evidence_quote":"Gives the Augmented Lagrangian methodology used in ALG 3 and ALG 4."}],"review_version":1}