{"id":"8debf64f-70be-4ea3-a15b-f17f6a2dbcec","arxiv_id":"1908.00996","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A nonlinear Petrov-Galerkin method that minimizes residuals in Lq-type norms is applied to convection-diffusion-reaction equations, numerically demonstrating that over- and undershoots near boundary layers vanish as q tends to 1 on suitable meshes.","lead":"This paper designs a finite element method for convection-diffusion problems that can suppress spurious oscillations near sharp layers by solving the equations in Lq spaces with q close to 1. The authors show numerically that on appropriately chosen meshes the overshoots and undershoots, the so-called Gibbs phenomena, vanish as q approaches 1.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Gibbs-elimination claim does not follow from the W^{1,q} quasi-optimality bound (3.12): for q close to 1 in 2D that norm does not control pointwise overshoot, and the method is not the Lq(Ω)-best approximation for ε>0.","rationale":"The reader correctly identified the unverified Fortin projector condition as a serious gap in the theory of the inexact method. My concern is distinct and, in my reading, more directly load-bearing for the headline claim: even if a Fortin projector with uniformly bounded norm exists and (3.12) holds, the resulting quasi-optimality is in W^{1,q}_0, a norm that does not control pointwise overshoot for q close to 1 in two dimensions. Moreover, the equivalence between the method and the Lq(Ω)-best approximation is only shown for the ε=0 optimal-test-norm setting, so the step from residual minimization to the Gibbs-behavior claim is not justified by the analysis in Section 3. The numerical evidence in Section 5 is substantial and genuinely interesting, and the q→1 behavior is demonstrated on several examples; my objection is to the strength of the wording 'Thus... can be eliminated', which presents an empirical regularity as a proven consequence. A conditional acceptance remains appropriate, but the authors should either prove a pointwise or L∞-type stability estimate, or explicitly state that the Gibbs-elimination claim is an observed numerical property supported by the experiments and by the heuristics of [5]. This does not change the reader's overall conditional verdict, so I recommend UNCHANGED.","tokens_in":22911,"tokens_out":8599,"duration_ms":91109,"concrete_test":"For Example 5.3 (corner-layer problem) on the modified Mesh 2, with q=1.01, ε=10^{-6}, and p_n=1, compute (i) the true Lq(Ω)-best approximation of the analytical solution in U_n and (ii) the solution un of the fully discrete method (4.8). Compare max(un−û) with max(best−û), where û is the interpolant in U_n, over a sequence of uniformly refined versions of the mesh. If the difference in the maximal overshoot does not tend to zero, or if it cannot be bounded in terms of the quasi-optimality constant C in (3.12), then the qualitative equivalence claim is not a consequence of the theory. A complementary analytical check: for d=2 and q<2, exhibit two functions w1,w2 in U_n with equal ‖u−w_i‖_{W^{1,q}} but with maximal overshoot differing by O(1), which would show that (3.12) alone cannot certify Gibbs elimination.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the discrete solution un qualitatively behaves like the Lq(Ω)-best approximation, so that q→1 removes Gibbs oscillations whenever the L1(Ω)-best approximation has no over- and undershoots. The only theoretical route to this conclusion is the quasi-optimality estimate (3.12), which requires the Fortin projector (3.11) and the inf-sup constants. That estimate is in the W^{1,q}_0 norm on U. For d=2 and q=1.01, W^{1,q}_0 embeds only into L^{q*} ≈ L^{2.02}, not into L∞; hence a small W^{1,q} error does not control the size of the pointwise over- and undershoots that define the Gibbs phenomenon. Separately, for ε>0 the objective minimized in (4.8) is the residual in the test norm (4.4), which is not the Lq(Ω) error; the exact identification with Lq(Ω)-best approximation is only established in the ε=0, optimal-test-norm case of Section 4.4.2. Thus even if a Fortin projector with bounded norm exists, (3.12) does not imply the qualitative Gibbs statement. The support for the central claim therefore rests on the numerical comparison in Section 5.3.1 (Figure 12) for one example and on the unpublished mesh-design predictions in reference [5]. The wording in the abstract and conclusions, 'Thus, the Gibbs phenomenon can be eliminated by taking the limit q→1', overstates what the analysis actually establishes.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a nonlinear Petrov-Galerkin method for the convection-diffusion-reaction equation in a W^{1,q}_0(Ω)-W^{1,q'}_0(Ω) Banach space setting, extending the DPG residual-minimization framework via duality mappings. A fully discrete inexact mixed method is derived, several test norms and weak boundary conditions are examined, and extensive one- and two-dimensional numerical experiments are presented. The central claim is that the discrete solution qualitatively behaves like the Lq(Ω)-best approximation, so that taking q→1 eliminates Gibbs phenomena whenever the L1(Ω)-best approximation has no overshoots; the paper demonstrates this on selected meshes, while honestly reporting meshes on which overshoots persist.","tokens_in":23250,"tokens_out":3764,"duration_ms":37945,"significance":"If the central claim were established, the paper would make a significant contribution to non-oscillatory finite element methods, connecting nonlinear Petrov-Galerkin ideas to L1-type best approximation and providing practical mesh-design guidance. The abstract framework with duality mappings is carefully presented, the numerical study is extensive and includes counterexamples (e.g., Mesh 3 in Section 5.3.1), and the comparison with Lq-best approximations in Figure 12 is informative. However, the theoretical route to the Gibbs-elimination claim is incomplete: the quasi-optimality estimate is in a norm that does not control pointwise overshoot, the identification with Lq-best approximation is only established in a zero-diffusion limit, and the fully discrete analysis relies on an unproven Fortin projector. The paper's own statements in Sections 3.3 and 5.2.3 acknowledge the latter difficulty.","major_comments":[{"comment":"The central claim that Gibbs phenomena can be eliminated by taking q→1 does not follow from the quasi-optimality estimate (3.12). That estimate controls the W^{1,q}_0(Ω) norm of the error, but for d=2 and q close to 1 the space W^{1,q}_0(Ω) does not embed into L∞(Ω), so a small W^{1,q} error does not control the size of the pointwise over- and undershoots that define the Gibbs phenomenon. Without an additional argument, the qualitative statement in the abstract and conclusions overstates what (3.12) establishes.","section":"Section 1.3, Conclusions, Eq. (3.12)"},{"comment":"For ε>0, the method (4.8) minimizes the residual in the test norm (4.4), which is not the Lq(Ω) error; the exact identification with the Lq(Ω)-best approximation is only derived in the ε=0, optimal-test-norm case of Section 4.4.2. Therefore, even if a bounded Fortin projector exists, the analysis does not show that the discrete solution is close to the Lq-best approximation for the convection-diffusion-reaction equation with ε>0. The numerical evidence in Figure 12 is suggestive but concerns a single example and does not replace a proof.","section":"Section 4.4.2 and Eq. (4.8)"},{"comment":"The error bound (3.12) requires the existence of a Fortin projector Π: V → V_m satisfying (3.11). The paper correctly states that dim(V_m) ≥ dim(U_n) is only necessary and that finding sufficient conditions is 'highly non-trivial'; the observation that compatibility 'is typically satisfied' for Δp ≥ 2 is numerical, not proven. If no such projector exists, the quasi-best approximation property of the fully discrete method does not follow, so this is a load-bearing gap in the theory.","section":"Eq. (3.11) and Section 5.2.3"},{"comment":"The inf-sup and continuity constants from the cited work [28] are not parameter-robust, and the paper acknowledges in Section 3.2 that extending robust estimates from Hilbert to Lq spaces remains an open problem. Consequently, even the abstract quasi-optimality route (3.12) does not provide the parameter-independent control that would be needed to justify the robustness claims in the convection-dominated regime; the observed robustness in ε is presently supported only by numerical experiments.","section":"Section 4, intro, and Section 3.2"}],"minor_comments":[{"comment":"Reference [5] is cited as an unpublished result, yet it carries a substantial part of the mesh-design predictions used in Section 5.3; it should be marked as 'in preparation' or 'personal communication' and, ideally, the relevant statements should be summarized in an appendix.","section":"References, [5]"},{"comment":"The captions of Figures 7 and 8 appear to be identical, both describing variation of ε, while the text of Section 5.2.3 attributes the h-refinement experiment to Figure 7; the captions need to be corrected.","section":"Figures 7 and 8"},{"comment":"The statement that for Δp ≥ 2 the Fortin condition 'is typically satisfied' is vague; specifying the meshes, polynomial degrees, and q-ranges for which this was observed would make the claim reproducible.","section":"Section 5.2.3"},{"comment":"The notation Δp is used without a definition; it should be defined as Δp = p_m − p_n when the spaces are introduced.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The paper's numerical program is interesting and the experiments are honest, including cases where the proposed Gibbs elimination fails. The main issue for publication is that the abstract and conclusions assert a theoretical connection to Lq-best approximation that the analysis does not establish. I would ask the authors to either prove the missing steps (Fortin projector existence, L∞ or pointwise control from the W^{1,q} estimate, and an ε>0 identification with Lq-best approximation) or substantially soften the claims and present the connection as a conjecture supported by numerical evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid incremental extension of the Banach-space residual-minimization framework to Lq for convection-diffusion-reaction, with extensive numerical experiments showing that q→1 can remove over- and undershoots on some meshes. The explanatory claim that the method behaves like the Lq(Ω)-best approximation is plausible, and the direct numerical comparison in Figure 12 supports it for one example. But the theoretical route to that claim is incomplete: the quasi-optimality bound (3.12) is in W^{1,q}, which for q close to 1 in 2D does not control pointwise overshoot, and exact Lq-best approximation is only established in the ε=0, optimal-test-norm case. So the Gibbs-elimination statement is an empirical observation with a heuristic explanation, not a theorem the paper proves.\n\nWhat the paper does well: the duality-mapping derivations are careful and correct; the inexact mixed formulation is clearly presented; and the numerical section is unusually honest, including a mesh (Mesh 3) where overshoots persist as q→1 and a modified-mesh case where they disappear. The link to the companion analysis of Lq-best approximation of discontinuities is reasonable, even if that work is unpublished.\n\nThe soft spots are real but not fatal if the paper is read as a numerical study. The Fortin projector condition (3.11) is never proven; it is only noted that dim(Vm) ≥ dim(Un) is necessary and that Δp≥2 seems to work numerically. The inf-sup constants from [28] are parameter-dependent, so the quasi-optimality constant may degrade with ε. Most importantly, the mesh-design predictions rely on the unpublished same-author result [5], and no code or data are provided. The abstract and conclusions overstate the result: 'the Gibbs phenomenon can be eliminated by taking the limit q→1' should be qualified as demonstrated for simple examples on carefully chosen meshes, with supporting analysis still pending.\n\nWho this is for: researchers working on DPG methods, Lq finite elements, and stabilized methods for convection-dominated problems. It deserves a serious referee: the framework generalization is meaningful, the experiments are informative, and the open theoretical questions are worth stating precisely. My recommendation: send it to peer review with the expectation of major revision, asking the authors to either prove or clearly scope the Gibbs-elimination claim, make the dependence on [5] explicit, and ideally release the code. I would cite it if I worked in this area.","headline":"A credible numerical study of an Lq variant of DPG-style residual minimization that shows q→1 suppresses Gibbs oscillations on some meshes, but the theoretical argument for the central claim is incomplete and the title oversells it.","tokens_in":23802,"tokens_out":1981,"would_cite":true,"duration_ms":22244,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N30","35J20"],"pacs":[],"model":"deepseek-v4-flash","headline":"A nonlinear Petrov-Galerkin method that minimizes residuals in Lq spaces makes numerical approximations of convection-dominated problems track the Lq-best approximation, so driving q toward 1 removes Gibbs oscillations on suitable meshes.","keywords":["convection-diffusion-reaction","Gibbs phenomenon","Petrov-Galerkin method","Banach spaces","duality mappings","Lq-best approximation","finite element method","residual minimization"],"falsifier":"A direct test: choose the one-dimensional non-uniform mesh for which the paper's companion analysis shows the L1-best approximation of a step keeps a positive overshoot as q→1, run the method with q→1, and compare min(un) with the L1-best interpolant's overshoot; if the undershoot does not approach the same positive limit, the claimed qualitative equivalence fails. Alternatively, search for a pair (Un, Vm) with dim Vm ≥ dim Un for which the Fortin condition fails and show the error bound is violated.","tokens_in":22675,"feed_emoji":"📉","tokens_out":8795,"duration_ms":83723,"temperature":0.7,"pith_summary":"This paper claims that a nonlinear Petrov-Galerkin method based on residual minimization in Lq-type Sobolev dual norms can remove the spurious oscillations that plague finite element solutions of convection-dominated convection-diffusion-reaction equations. The authors extend the discontinuous-Petrov-Galerkin idea from Hilbert spaces to reflexive Banach spaces by replacing the Riesz map with a duality mapping, and they show that the resulting approximations track the Lq(Ω)-best approximation of the exact solution. As q tends to 1, Gibbs phenomena disappear on meshes where the L1-best approximation of the underlying layer is oscillation-free, which they verify in one and two dimensions. The result matters because it offers a parameter-controlled route to oscillation-free layer approximation without adding artificial diffusion.","feed_headline":"A q-to-1 residual-minimization scheme kills Gibbs oscillations","feed_subtitle":"On meshes where the L1 best approximation is oscillation-free, the computed solution follows suit as q approaches 1.","key_machinery":"The load-bearing object is the duality mapping JVφ: V → V', the Banach-space replacement for the Riesz map, defined for a weight φ by pairing with norm-bounded functionals. With V = W1,q'0(Ω) and φ(t) = tq'-1, the mixed system—find un ∈ Un and residual r ∈ Vm such that ⟨$JVφ^{{-1}}$(r), vm⟩ + Bε(un, vm) = ℓ(vm) for all vm and Bε(wn, r) = 0 for all wn—turns residual minimization in the dual norm into a finite-dimensional nonlinear saddle-point problem. The parameter q appears only through this mapping, so it directly selects the norm in which quasi-best approximation is measured; q→1 forces the near-L1 behavior that suppresses oscillations, while the unproven Fortin-projector condition is what would guarantee that the discrete solution really is quasi-best.","core_discovery":"The paper's central claim is that the numerical solution produced by its nonlinear Petrov-Galerkin method behaves, up to the quasi-best approximation constants, like the Lq(Ω)-best approximation of the exact solution. Consequently, when q is driven toward 1, Gibbs-type over- and undershoots near boundary and interior layers disappear on exactly those meshes where the L1(Ω)-best approximation of the layer is itself oscillation-free. The authors demonstrate this by minimizing the residual in the dual norm of W1,q'0(Ω) with the duality mapping of weight tq'-1, and they show in one- and two-dimensional convection-dominated examples that their computed solutions match the overshoot profile of the true Lq best approximation; on meshes known to produce persistent L1 overshoots, the oscillations persist, and on suitably modified meshes they vanish as q→1.","pith_inferences":["If the qualitative match with Lq best approximation holds beyond the tested examples, the method could be used as a numerical probe for the Lq-best approximation of discontinuous solutions themselves, giving a PDE-based way to explore mesh-dependent Gibbs phenomena.","The unproven Fortin condition suggests a concrete research target: deriving sufficient compatibility conditions for the discrete trial and test spaces, or importing an adaptive strategy that bypasses discrete inf-sup conditions, would turn the method from a numerically observed recipe into a guaranteed one.","Because monotone high-order schemes for nonlinear conservation laws are necessarily nonlinear, this nonlinear framework may extend to shock problems where q→1 acts as a built-in limiter, though the computational cost of solving the nonlinear system with q' large would have to be addressed first."],"forward_implications":["For convection-dominated problems, choosing q close to 1 yields approximations without over- and undershoots on meshes where the L1-best approximation of the layer is oscillation-free; the layer does not need to be fully resolved.","The method's convergence rate in the convection-dominated regime is roughly O(h^{1/q}), consistent with best approximation of a discontinuity, meaning the oscillation control comes at the expected cost of reduced order near layers.","Enriching the test space by two polynomial degrees and imposing weak inflow boundary conditions on the residual variable is enough, in the tested cases, to obtain the quasi-optimal behavior; a non-constant weighting function in the test norm is unnecessary.","The observed match with Lq best approximation gives a practical mesh-design rule: refine elements near the layer and align the closest interior nodes parallel to it; overshoots then vanish as q→1, as in the corner-layer and interior-layer examples.","On meshes where the L1-best approximation provably keeps overshoots, the method also keeps overshoots as q→1, so the Gibbs phenomenon is not eliminated universally but only where the underlying approximation space allows it."],"supporting_citations":[{"why":"Supplies the abstract residual-minimization and nonlinear mixed method in Banach spaces that the paper instantiates.","marker":"[17]"},{"why":"Provides the mesh-dependent Lq-best-approximation overshoot analysis used to predict when q→1 eliminates Gibbs phenomena.","marker":"[5]"},{"why":"Supplies the H1-setting dual Petrov-Galerkin method for convection-diffusion and the idea of weak inflow boundary conditions.","marker":"[19]"},{"why":"Gives the earlier Lp finite element residual-minimization approach to first-order PDEs that the limit case is compared with.","marker":"[7]"},{"why":"Establishes the inf-sup condition for the bilinear form in the W1,q-W1,q' setting needed for well-posedness.","marker":"[28]"},{"why":"Analyzes the discrete-dual minimal-residual method for weak advection-reaction problems and provides special cases for Fortin compatibility.","marker":"[40]"},{"why":"Introduces the Fortin-operator analysis for practical DPG methods that underlies the quasi-optimality estimate used here.","marker":"[38]"}],"fun_headline_variants":["As q→1, Petrov-Galerkin erases Gibbs oscillations","Nonlinear Petrov-Galerkin: no Gibbs as q heads to 1","q→1 residual minimization kills boundary-layer oscillations","Oscillation-free layers when q approaches 1","Lq-best approximation: key to killing Gibbs overshoots"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The conclusion that the computed solution is a quasi-best Lq approximation rests on the existence of a Fortin projector between the discrete test and trial spaces that the paper never proves; it only observes numerically that enriching the test space by two polynomial degrees seems to work.","fun_headline_variants_meta":{"raw":{"variants":["As q→1, Petrov-Galerkin erases Gibbs oscillations","Nonlinear Petrov-Galerkin: no Gibbs as q heads to 1","q→1 residual minimization kills boundary-layer oscillations","Oscillation-free layers when q approaches 1","Lq-best approximation: key to killing Gibbs overshoots"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000907,"raw_usage":{"total_tokens":3924,"prompt_tokens":990,"completion_tokens":2934,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":2847}},"tokens_in":606,"tokens_out":2934,"duration_ms":19103,"temperature":1.0,"reasoning_tokens":2847,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:26:04.279817+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test: choose the one-dimensional non-uniform mesh for which the paper's companion analysis shows the L1-best approximation of a step keeps a positive overshoot as q→1, run the method with q→1, and compare min(un) with the L1-best interpolant's overshoot; if the undershoot does not approach the same positive limit, the claimed qualitative equivalence fails. Alternatively, search for a pair (Un, Vm) with dim Vm ≥ dim Un for which the Fortin condition fails and show the error bound is violated.","supporting_citations":[{"cited_title":"Houston, S","cited_arxiv_id":null,"evidence_quote":"Provides the mesh-dependent Lq-best-approximation overshoot analysis used to predict when q→1 eliminates Gibbs phenomena."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the H1-setting dual Petrov-Galerkin method for convection-diffusion and the idea of weak inflow boundary conditions."},{"cited_title":"Guermond, A finite element technique for solving first-order PDEs in LP , SIAM J","cited_arxiv_id":null,"evidence_quote":"Gives the earlier Lp finite element residual-minimization approach to first-order PDEs that the limit case is compared with."},{"cited_title":"Houston, I","cited_arxiv_id":null,"evidence_quote":"Establishes the inf-sup condition for the bilinear form in the W1,q-W1,q' setting needed for well-posedness."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Analyzes the discrete-dual minimal-residual method for weak advection-reaction problems and provides special cases for Fortin compatibility."},{"cited_title":"Gopalakrishnan, W","cited_arxiv_id":null,"evidence_quote":"Introduces the Fortin-operator analysis for practical DPG methods that underlies the quasi-optimality estimate used here."}],"review_version":1}