{"id":"b65e0abb-6d55-49ae-b642-f881baf9655c","arxiv_id":"2412.01232","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The authors extend their prior dual (Lagrange multiplier) variational framework to steady and transient convection-diffusion and heat equations, discretizing with B-splines or RePU networks, and report numerical accuracy with observed convergence rates.","lead":"This paper applies a dual variational principle to convection-diffusion and heat equations, producing symmetric Galerkin systems that need no upwinding or stabilization. The method is tested with B-splines and fixed-weight RePU neural networks, and the authors report accurate solutions and observed convergence rates.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Transient accuracy claim rests on an undemonstrated buffer-zone remedy for the terminal-time boundary layer caused by the arbitrary dual condition λ(x,1)=0; this is the load-bearing weak spot.","rationale":"We agree with the reader that the terminal boundary layer is the weakest assumption. The steady-state results (Figs. 12–13) are credible and the uniqueness proof in §4.1 is sound, so the dual formulation itself is not in question. The convergence-rate language in the abstract is also overstated, as the rates are observed numerically and no stability estimate is proved; however, that overstatement affects the framing more than the numerical method. The transient case is where the method's novel space-time capability is demonstrated, and the error concentration near t=1 is visible in Fig. 16. Since the paper's own remedy is cited to prior work but not demonstrated, a conditional verdict is appropriate. We would not change the reader's verdict.","tokens_in":35245,"tokens_out":10409,"duration_ms":97402,"concrete_test":"Implement the buffer-zone remedy for the §5.3 example: set T=1, δ=0.1, solve the dual problem with λ(x,1+δ)=0 and λ(1,t)=0 on [0,1+δ] using the same B-spline degrees p=9, q=10 (or the same number of DOF per unit interval), with the initial condition and boundary data extended arbitrarily on [1,1+δ]. Then recompute the maximum relative L2 error in u and q over the interior slice [0,1]×[0,1] and compare with the reported 0.06/0.1. If the errors drop below 1%, the remedy works and the conditional verdict stands; if they remain above 5%, the §5.3.1 explanation is incomplete and the transient accuracy claim needs to be downgraded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's strongest numerical claim for a genuinely time-dependent problem is the transient convection-diffusion result in §5.3, where the reported maximum relative errors are 0.06 for u and 0.1 for q, concentrated near t=1 (§5.3.1). Section 5.3.1 explains that the arbitrary choice λ(x,1)=0, combined with the boundary condition λ(1,t)=0, produces a corner conflict at (1,1): compatibility with the primal solution forces ∂tλ(1,1) and ∂xμ(1,1) to values that a smooth global polynomial cannot represent, creating a dual boundary layer. The paper's proposed remedy — extending the time domain to [0,T+δ] and discarding the strip (T,T+δ] — is described but never implemented or tested here. The central claim that the method 'delivers sound accuracy for ODEs and PDEs' therefore depends on an untested assumption: that this buffer-zone truncation actually removes the contamination in the interior [0,T]. If the remedy fails in more demanding regimes (smaller κ, longer times, or sharper data), the transient space-time formulation loses its accuracy guarantee near the terminal time, weakening the paper's most novel contribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a duality-based variational formulation for PDEs that lack a primal variational structure, treating the PDE as a constraint and optimizing a user-chosen convex auxiliary potential. For the one-dimensional transient convection-diffusion equation, the authors derive a dual functional whose first variation is claimed to reproduce the primal weak form after the dual-to-primal map is substituted. They discretize the resulting symmetric variational problem with shallow RePU neural networks and B-splines in steady state, and with tensor-product B-splines in a space-time setting for transient problems. Numerical examples include Laplace's equation, steady-state and transient convection-diffusion, and transient heat conduction, together with a uniqueness proof for the dual variational system and numerical convergence studies for the steady-state problem.","tokens_in":35408,"tokens_out":9673,"duration_ms":91373,"significance":"If its claims hold, the paper offers a systematic way to obtain symmetric Galerkin formulations for non-variational PDEs without stabilization parameters, and it extends the dual approach to space-time discretizations using smooth approximants. The explicit stiffness-matrix derivations in Section 4, the uniqueness proof in Section 4.1, and the exact recovery of the Laplace solution in Section 5.1 are genuine strengths, as are the bracketing of the arbitrary terminal dual condition in the ODE example and the numerical convergence rates for the steady-state convection-diffusion problem. However, the transient results, which are the most novel contribution, currently carry an acknowledged terminal-time boundary-layer error whose proposed remedy is not implemented or tested; this weakens the paper's broad accuracy claim until the issue is resolved.","major_comments":[{"comment":"The main time-dependent demonstration reports relative maximum errors of 0.06 for u and 0.1 for q, with the errors explicitly concentrated near the terminal time t=1. The paper attributes this to the corner conflict caused by the arbitrarily chosen terminal condition λ(x,1)=0 and then proposes a buffer-zone extension of the time domain to [0,T+δ], discarding (T,T+δ], but this remedy is never implemented or tested. Moreover, no h- or p-refinement study is reported for the transient convection-diffusion problem, so it is not demonstrated that the contamination is localized and decreases with resolution. Because the abstract claims that the method 'delivers sound accuracy for ODEs and PDEs,' this transient accuracy claim is load-bearing and currently rests on an untested assumption. Please implement the buffer-zone or time-slicing strategy and report errors on the original interval [0,T], or provide a refinement study for the space-time discretization showing that the terminal-layer error is confined and reduced under refinement.","section":"Section 5.3 and 5.3.1"},{"comment":"The central theoretical step is the statement that the vanishing first variation of the dual functional S[λ,μ] is, up to Dirichlet conditions on λ, the weak form of the primal IBVP with the dual-to-primal map incorporated. The paper delegates this to 'familiar lines' and does not show the explicit integration-by-parts calculation. Since this equivalence is the foundation of the numerical discretization in Section 4 and also clarifies why λ may be assigned a terminal condition while u cannot, please include the full derivation, showing how the boundary data u1,u2 and the initial condition u0 arise as natural conditions from (53)-(54). This will also make precise the sense in which the top-edge dual boundary condition creates the terminal boundary layer discussed in Section 5.3.1.","section":"Section 3.2, Eq. (46)"}],"minor_comments":[{"comment":"The tuple D in the transient heat equation should contain ∂μ/∂x, not ∂u/∂x; as written, it is inconsistent with the dual-to-primal map used in (52a).","section":"Eq. (52a)"},{"comment":"The abstract says rates of convergence are 'established,' but the paper provides numerically observed slopes and heuristic interpolation estimates rather than a convergence theorem; please qualify the claim as numerical or provide a proof.","section":"Abstract and Section 5.2.2"},{"comment":"The RePU 'neural network' approximants have fixed hidden-layer weights and biases, so only the output-layer coefficients are unknown and the approximation is linear in the unknowns; this should be stated explicitly at the first use to avoid overclaiming a machine-learning training procedure.","section":"Section 5.2.1"},{"comment":"The integral notation ∫_0^1∫_t^1 ... is confusing because t is used both as the time variable and as the lower limit of integration; replacing the inner integration variable with τ would improve readability.","section":"Section 4.1"},{"comment":"For α=50 with p=q=1, the reported rates (1.2, 0.9) are not close to the predicted order one; a sentence explaining the effect of the boundary layer on the coarsest discretizations would help.","section":"Figure 13"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the underlying dual variational idea is coherent; the derivations of the stiffness matrices and the uniqueness argument are sound to my reading. The main risk is the transient accuracy claim, which depends on an unverified buffer-zone remedy for the terminal-time boundary layer. If the authors can demonstrate the remedy or otherwise show controlled terminal-time errors, the paper would be acceptable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the quick take: this is a solid numerical follow-up in the authors' dual variational program, with a genuinely new weak form for the transient convection-diffusion equation and a correct uniqueness proof. The abstract oversells the convergence-rate claim, and the proposed fix for the terminal-time boundary layer is untested. Neither flaw sinks the paper, but they should be addressed before publication.\n\nWhat's actually new: Section 3.2 derives the dual weak form for the transient convection-diffusion equation by treating (36) as a constraint and using a quadratic auxiliary potential. The resulting symmetric stiffness matrix in Section 4 is derived cleanly, and the uniqueness proof in Section 4.1 is internally consistent: the bilinear form reduces to a sum of squares, and the energy argument forces λd and μd to zero. That is real, checkable mathematics. The numerical work is also honest: the steady-state convergence study shows clean rates of roughly p and p+1 in L2 and H1 for B-splines, and the RePU results are consistent even though the basis is a frame rather than a basis. The authors also give a candid explanation of the corner conflict at (1,1) in Section 5.3.1, which is more than most papers do.\n\nThe soft spots: first, the abstract claims rates are \"established,\" but the paper only reports observed slopes from log-log plots and invokes standard interpolation estimates. There is no stability proof for the discrete Galerkin system; the uniqueness proof covers the continuous variational form, not the discrete system's convergence rate. Second, the terminal-time boundary layer is real: the arbitrary choice λ(x,T)=0 creates a corner conflict that produces 6-10% errors near t=1 in the transient example. The proposed buffer-zone fix is described but never implemented or tested, so the transient accuracy claim rests on an untested assumption. This is the load-bearing weak spot, though the authors flag it themselves. Third, no code is provided, which makes reproducing the B-spline computations more work than necessary.\n\nThe paper is worth a serious referee. It is a legitimate incremental step, the derivations check out, and the numerical experiments are transparent. I'd ask the referee to require a tempering of the abstract's \"established\" language, and to either test the buffer-zone idea or explicitly limit the transient claims to the tested range.","headline":"A solid incremental contribution to the dual variational program; the transient weak form and uniqueness proof are real, but the terminal-time accuracy claim is weakened by an untested remedy and the convergence rates are observed, not established.","tokens_in":36009,"tokens_out":2946,"would_cite":false,"duration_ms":27045,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N30","65M60","65N12"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper shows that PDEs without an exact primal variational structure, such as the transient convection-diffusion equation, can be solved through a dual variational principle: maximizing a convex functional in Lagrange-multiplier fields…","keywords":["dual variational principles","convex duality","weak formulation","space-time Galerkin method","B-splines","RePU neural networks","convection-diffusion equation","terminal boundary condition"],"falsifier":"Solve the transient convection-diffusion test with $\\kappa=0.01$, $\\alpha=0.1$, and $u_0=\\sin(2\\pi x)$ on uniformly refined space-time B-spline meshes, and measure the $L^2$ error on the terminal strip $[0,1]\\times[0.9,1]$; if that error does not shrink under refinement, or if appending a buffer zone $[0,1+\\delta]$ does not eliminate it, the assumption that arbitrary dual terminal data are harmless fails.","tokens_in":34955,"feed_emoji":"📐","tokens_out":11773,"duration_ms":96331,"temperature":0.7,"pith_summary":"This paper claims that PDEs lacking an exact primal variational structure, such as the transient convection-diffusion equation, can be solved through a dual variational principle. The PDE is treated as a constraint on an arbitrarily chosen convex auxiliary potential; stationarity with respect to the primal fields yields a dual-to-primal map, and the remaining problem is a concave maximization of a dual functional under Dirichlet conditions on the dual fields. The first variation of that dual functional reproduces the weak form of the original PDE with the dual-to-primal map substituted, so a standard Galerkin discretization applies without upwinding or stabilization parameters. The paper derives the dual weak form for the one-dimensional transient convection-diffusion equation, computes with RePU neural networks and B-splines, and reports convergence rates of order $p$ in the $L^2$ norm and $p+1$ in the $H^1$ seminorm for the steady case.","feed_headline":"Dual principle solves PDEs that lack a variational form","feed_subtitle":"Dual fields carry convection; no stabilization parameters, and convergence rates match theory.","key_machinery":"The load-bearing object is the dual functional $S[\\lambda,\\mu] = -\\frac{1}{2}\\int_0^1\\int_0^1\\big((u^H)^2+(q^H)^2\\big)\\,dx\\,dt + \\int_0^1(\\bar{u}_2\\mu(1,t)-\\bar{u}_1\\mu(0,t))\\,dt - \\int_0^1 u_0(x)\\lambda(x,0)\\,dx$, with $u^H$ and $q^H$ given by the dual-to-primal map. It is built by writing the PDE system as constraints in a Lagrangian, applying the stationarity conditions $\\partial L/\\partial u=0$ and $\\partial L/\\partial q=0$ to eliminate the primal fields, and then maximizing the resulting concave functional. Maximizing $S$ over admissible $\\lambda$ and $\\mu$ yields variational equations that are exactly the primal PDE together with its initial and boundary data, which is why the method turns non-variational problems into variational ones.","core_discovery":"The central discovery is that the vanishing of the first variation of the dual functional is, up to Dirichlet boundary conditions on dual fields, the weak form of the primal PDE problem with the dual-to-primal change of variables incorporated. For the transient convection-diffusion equation with diffusion $\\kappa$ and convection $\\alpha$, the dual fields $(\\lambda,\\mu)$ generate the primal fields through $u^H=\\partial_t\\lambda+\\partial_x\\mu$ and $q^H=\\mu-\\alpha\\lambda-\\kappa\\partial_x\\lambda$, and maximizing the quadratic dual functional over $\\lambda$ with homogeneous Dirichlet data on the sides and terminal time, and over free $\\mu$, delivers the primal solution. The discrete Galerkin system is symmetric, uniqueness of the dual solution is proven, and numerical experiments show clean convergence for steady convection-diffusion with rates matching those of the underlying B-spline degrees.","pith_inferences":["Editorial inference: if the terminal boundary-layer errors are as localized as the examples suggest, the buffer-zone and time-slicing remedies make the space-time dual method practical for long-time simulation, but the paper does not demonstrate that combination with a convergence study.","Editorial inference: the method's tolerance of linearly dependent RePU trial functions implies that overcomplete dictionaries can be used harmlessly as dual ansatz spaces, which could simplify constructing admissible neural-network trial spaces beyond standard basis constraints.","Editorial inference: the same constraint-plus-convex-potential construction could turn parameter identification or inverse problems into convex dual optimization, since the primal equation enters only as a constraint; this direction is not explored here."],"forward_implications":["A plain Galerkin discretization of the symmetric dual system works for convection-dominated regimes where the primal Galerkin form oscillates, so no upwinding or stabilization parameter is needed.","Initial-value problems are reformulated as boundary-value problems in space-time with a terminal Dirichlet condition on the dual field; for the ODE $\\dot u = au$ the paper proves the arbitrary terminal data do not change the primal solution.","For steady convection-diffusion, the numerical rates are approximately $p$ in the $L^2$ norm for $u$ and $p+1$ in the $H^1$ seminorm for $u'$ when the dual field $\\mu$ has degree $p$ and $\\lambda$ has degree $p+1$.","Because the highest-order derivatives in the dual problem belong to heat or linear transport, the corresponding dual boundary-value problem is degenerate elliptic yet has a unique solution, covering the pure-transport limit.","Errors that concentrate near the terminal time $t=1$ are attributed to a corner conflict between dual Dirichlet data and can, per the paper, be removed by extending the time interval with a buffer zone or by time-slicing."],"supporting_citations":[{"why":"introduces the dual variational scheme this paper applies: PDEs as constraints with an auxiliary convex potential.","marker":"[9]"},{"why":"supplies the formal hidden-convexity justification that the dual variational equations are locally degenerate elliptic.","marker":"[11]"},{"why":"gives rigorous analysis and prior computational demonstration of the dual construction for nonlinear elasticity.","marker":"[12]"},{"why":"extends the dual variational framework to incompressible fluids, the motivating class of primal non-variational PDEs.","marker":"[13]"},{"why":"provides prior computational implementation for heat, linear transport, and rigid-body rotation, including time-slicing and the terminal boundary-layer remedy.","marker":"[17]"},{"why":"solves inviscid Burgers as a degenerate elliptic problem, supporting the pure-transport limit and the buffer-zone strategy.","marker":"[18]"}],"fun_headline_variants":["Dual trick solves PDEs without a primal variational form","B-splines and neural nets solve PDEs via dual principle","Dual fields make PDEs variational without primal structure","Symmetric Galerkin from dual principle for convection-diffusion","Dual principle turns non-variational PDEs into variational ones"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the arbitrarily chosen Dirichlet data on the dual field, especially the terminal condition $\\lambda(x,1)=0$, do not contaminate the computed primal solution even though they create a corner conflict and boundary layers near $t=1$; if those layers pollute the interior, the transient accuracy claims weaken.","fun_headline_variants_meta":{"raw":{"variants":["Dual trick solves PDEs without a primal variational form","B-splines and neural nets solve PDEs via dual principle","Dual fields make PDEs variational without primal structure","Symmetric Galerkin from dual principle for convection-diffusion","Dual principle turns non-variational PDEs into variational ones"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001318,"raw_usage":{"total_tokens":5426,"prompt_tokens":1059,"completion_tokens":4367,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":675,"completion_tokens_details":{"reasoning_tokens":4281}},"tokens_in":675,"tokens_out":4367,"duration_ms":27743,"temperature":1.0,"reasoning_tokens":4281,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:33:46.942253+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Solve the transient convection-diffusion test with $\\kappa=0.01$, $\\alpha=0.1$, and $u_0=\\sin(2\\pi x)$ on uniformly refined space-time B-spline meshes, and measure the $L^2$ error on the terminal strip $[0,1]\\times[0.9,1]$; if that error does not shrink under refinement, or if appending a buffer zone $[0,1+\\delta]$ does not eliminate it, the assumption that arbitrary dual terminal data are harmless fails.","supporting_citations":[{"cited_title":"Acharya, Variational principle for nonlinear PDE systems via duality, Quarterly of Applied Mathematics 81 (2023) 127–140","cited_arxiv_id":null,"evidence_quote":"introduces the dual variational scheme this paper applies: PDEs as constraints with an auxiliary convex potential."},{"cited_title":"Singh, J","cited_arxiv_id":null,"evidence_quote":"gives rigorous analysis and prior computational demonstration of the dual construction for nonlinear elasticity."},{"cited_title":"Kouskiya, A","cited_arxiv_id":null,"evidence_quote":"provides prior computational implementation for heat, linear transport, and rigid-body rotation, including time-slicing and the terminal boundary-layer remedy."}],"review_version":1}