{"id":"0781d07a-0347-485b-bea1-c2dc3bc5f6fa","arxiv_id":"1908.08981","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"An ultraweak variational formulation with DPG approximation is developed for nondivergence-form elliptic PDEs with Cordes coefficients, with well-posedness, quasi-optimality, Fortin operators, and a posteriori estimators.","lead":"This paper builds a new computer-friendly formulation for a stubborn class of equations, the 'nondivergence form' PDEs that appear in optimal control, and proves the resulting method converges. It also supplies automatic error indicators so the computer can refine the mesh where the solution is rough, with 2D experiments included.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fully discrete DPG convergence is proven only for A∈P^p(T); the Section 6.3 radial-annulus example falls outside Theorem 23, so the general-Cordes DPG claim rests on an unproven Fortin assumption.","rationale":"The reader's weakest assumption correctly identifies the Fortin gap for general Cordes coefficients, and this is the most load-bearing limitation of the paper as written: the fully discrete DPG convergence theorems and a posteriori bounds require a Fortin operator that Theorem 23 supplies only for piecewise-polynomial A. However, the reader's formulation is slightly too broad, because the DPG-LS scheme of Section 2.6 is fully discrete and provably convergent for arbitrary Cordes A: its Fortin condition (17) involves only c(·,·), which does not contain A, and Lemma 21 gives the required Π^divDiv. Moreover, Section 2.7 already hints at the natural fix Vh=A(Mh)×Qh, which would extend the DPG theory to any bounded measurable A, so the gap is a limitation of the chosen test space and of the paper's presentation rather than a fundamental flaw in the method. The numerical experiments, especially Section 6.3 where the authors acknowledge the non-equivalence, are consistent with this reading. No internal inconsistency or unsupported central theorem was found in the well-posedness proof; the main theorems are supported by written arguments, with Lemma 11 cited from the author's published work. The verdict CONDITIONAL remains appropriate: the paper should either prove the Fortin property for the natural A-dependent test space, or explicitly restrict the fully discrete DPG convergence claims to the piecewise-polynomial case while crediting the proven DPG-LS variant for general Cordes coefficients.","tokens_in":15418,"tokens_out":15681,"duration_ms":174117,"concrete_test":"Re-run Example 6.3 with the DPG method using the test space Vh=A(Mh)×Qh (first component generated by A times each P0 basis function) instead of Vh^0, and on the same meshes compute the discrete inf-sup constant β_h = inf_{0≠uh∈Uh} sup_{0≠vh∈Vh} b(uh,vh)/(||uh||_U ||vh||_V) for both choices. If the exact-space β_h stays bounded below while the Vh^0 β_h decays under refinement, the missing Fortin assumption is load-bearing for the implemented method; if both stay bounded, the paper's Vh^0 scheme is numerically stable and the gap is a proof-technicality that should be stated as such.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper proves well-posedness of the ultraweak formulation (Theorem 4) and, for the DPG-LS scheme of Section 2.6, fully discrete quasi-optimality under the Fortin condition (17); because the form c in (14) is independent of A, Lemma 21 supplies a Fortin operator for arbitrary Cordes A, so that variant is covered. The load-bearing gap is the DPG method of Section 2.5 as implemented: Theorem 23 constructs a Fortin operator only for A∈P^p(T) with Vh^p=P^p×Qh, and Lemma 21's Π^divDiv alone cannot fix the scalar component unless A:Mh lies in the chosen polynomial space. Section 6.3 deliberately uses coefficients whose discontinuities are not aligned with the mesh, so A∉P^p(T) on the active triangulations; the paper itself states that the two discrete methods are then no longer equivalent and Theorem 23 does not hold. Consequently Theorems 7, the a posteriori equivalence (22), and Theorem 14 do not, as written, apply to the DPG method in the general Cordes regime advertised by the abstract; the numerical results there are evidence, not proof. The gap is not a fundamental obstruction: taking the first test component as Nh=A(Mh) (Section 2.7) and ΠF=(Π_Nh,Π^divDiv) would satisfy (10) for any bounded A, so the concern is about what the paper proves versus what it could prove, and the verdict should remain conditional.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops an ultraweak variational formulation for linear elliptic PDEs in nondivergence form with Cordes coefficients. The formulation introduces the Hessian M=D^2u and a broken trace variable, leading to a well-posed problem in U=L2(Ω)×L2_sym(Ω)×Uhat. The paper proves equivalence with the strong problem, proposes DPG and DPG-least-squares discretizations, proves quasi-optimality under a Fortin assumption, constructs Fortin operators for lowest-order spaces, gives a posteriori error estimators with reliability and efficiency, and reports numerical experiments with uniform and adaptive refinement.","tokens_in":15732,"tokens_out":10313,"duration_ms":110820,"significance":"If the central claims hold, the paper is a useful contribution: it brings the DPG framework to Cordes-coefficient nondivergence-form problems, avoids H2-conforming elements by using only traces of rHCT elements, produces symmetric positive definite systems, and supplies local error indicators. The analysis is largely self-contained and the proofs of well-posedness and quasi-optimality are written out; the Fortin operator for the DPG-least-squares variant works for arbitrary Cordes coefficients because the form c in (14) is independent of A. The main weakness is a mismatch between the general Cordes-coefficient scope advertised in the abstract and the fully discrete convergence theory for the DPG method of Section 2.5 as implemented: that theory requires A to be piecewise polynomial on the mesh. This is a scope issue rather than a fundamental obstruction, and it can be repaired either by narrowing the claims or by proving a Fortin operator for general A (e.g., using the space N_h=A(M_h) suggested in Section 2.7).","major_comments":[{"comment":"The fully discrete convergence result for the DPG method of Section 2.5, Theorem 7, and the a posteriori equivalence (22) of Theorem 14 are conditional on the Fortin assumption (10). The only construction for the implemented method is Theorem 23, which assumes A∈P^p(T) for the chosen polynomial test space V_h^p=P^p×Q_h. Section 6.3 uses coefficients discontinuous along circles that are not aligned with the mesh, so A∉P^p(T) on the active triangulations; the paper itself states that the two methods are then no longer equivalent and that Theorem 23 does not hold. Consequently, Theorems 7 and 14 do not, as written, apply to the DPG method for the general Cordes regime advertised in the abstract, and the Section 6.3 DPG results are numerical evidence rather than proven convergence. Please either restrict the abstract and introduction to the case covered by Theorem 23, or add a Fortin construction for general Cordes coefficients (for instance by taking the first test component to be N_h=A(M_h) as in Section 2.7 and verifying (10) with Π_F=(Π_N_h, Π_divDiv)); the latter is straightforward in principle but is not carried out in the paper.","section":"Sections 2.5, 5.2, 6.3; Theorems 7 and 14"},{"comment":"Lemma 11 is load-bearing for the proof of Theorem 4 because Proposition 13 uses it to identify the space V_0={v∈V : bhat(uh,v)=0 for all uh∈Uhat}. The proof is omitted and attributed to the author's own [14, Proposition 3.8] and [15, Proposition 11]. Since the cited statement is for traces of H^2_0(Ω) functions and the present setting uses traces of X=H^2(Ω)∩H^1_0(Ω), the transfer is not literally automatic. Please include the short adaptation argument or state precisely why the proof in [14] covers the present case; this will remove a nontrivial dependency from the central well-posedness theorem.","section":"Section 3.2, Lemma 11 and Proposition 13"}],"minor_comments":[{"comment":"In the proof of Theorem 14, the identity (f,v−Π_L2 v)=(f−A:Q_h,v−Π_L2 v) is asserted for any Q_h∈L2_sym(Ω). The Fortin property (10) supplies this only for Q_h belonging to the discrete space M_h (the second component of U_h), since it is verified for trial functions of the form (0,M_h,0). The subsequent choice Q_h=M_h is admissible, so the argument is repairable, but the sentence as written overstates the range of Q_h and should be corrected.","section":"Proof of Theorem 14, Section 4"},{"comment":"The text uses the notation 'Vh0' for the discrete test space; for consistency with Section 5.2 this should be V_h^0, and the polynomial degree p in V_h^p should be stated explicitly where the test spaces are first used in the experiments.","section":"Section 6, numerical examples"},{"comment":"The remark that a basis for N_h={A:M_h} is hard to determine is important for practice, but it would be helpful to add a sentence explaining that the theoretical Fortin construction with N_h is nonetheless valid for any bounded A, so that the limitation is one of implementation rather than of the ultraweak formulation itself.","section":"Section 2.7"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid contribution and the central well-posedness and semi-discrete analysis are internally consistent. My recommendation of major revision is driven by the claim-scope mismatch for the fully discrete DPG method in the general Cordes case, and by the omitted proof of a load-bearing trace lemma. The former is likely fixable by a modest theoretical addition, and the latter by a short argument; neither appears to be a fundamental error. The overlap with the author's previous work [12,14,15] is substantial, but the application to nondivergence-form equations with Cordes coefficients is new and the numerical study is informative."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nI've read Führer's ultraweak-DPG paper for nondivergence-form Cordes problems (1908.08981). Bottom line: the reader's conditional verdict is fair, and the main caveat to keep in mind is narrower than the abstract implies: the fully discrete DPG scheme is proven only for piecewise-polynomial coefficients; the DPG–least-squares variant covers the general case.\n\nWhat is new: this is the first ultraweak DPG treatment of Cordes-type nondivergence-form equations. The continuous formulation is well-posed (Theorem 4), equivalent to the strong problem (Prop 5), and the proof via breaking spaces is written out and internally consistent—I checked the sign conventions in (5) and Lemma 10; they work. The practical selling point is real: only traces of H^2 functions are needed, not H^2-conforming elements, so the rHCT trace space is a genuine simplification over Gallistl's LS-FEM. The a posteriori theory is standard DPG machinery, applied correctly.\n\nCredit where due: the paper is honest about its own limits—it says outright in Section 6.3 that Theorem 23 does not hold when coefficient discontinuities cut through elements, and it says \"we omit the proof\" of Lemma 11. That transparency makes the soft spots easier to weigh.\n\nSoft spots, in proportion:\n\n1. Lemma 11 (the trace characterization identifying V0) is load-bearing for well-posedness and is cited from the author's own [14] without proof. Proposition 3 is in the same boat. A referee should ask for a proof or a precise statement of the needed modification. This weakens self-containedness, not correctness.\n\n2. The bigger gap: Theorem 23 constructs the Fortin operator for the Section 2.5 DPG method only when A is piecewise polynomial of degree p. For general Cordes A—the regime the abstract advertises—that scheme's convergence theory is unproven. The DPG-LS variant of Section 2.6 is fully covered for any bounded A, since c in (14) is A-independent and Lemma 21 supplies the Fortin operator. So the abstract oversells slightly, but it is a scheme-specific gap, not a broken theory.\n\n3. The gap is patchable: taking the scalar test component as Nh = A(Mh), as Section 2.7 sketches, gives a Fortin operator for any bounded A. The stress-test's phrasing is right: the issue is what the paper proves versus what it could prove.\n\nMinor: no code or data shipped. Small deduction only.\n\nWho should read it: numerical analysts working on DPG methods or on nondivergence-form/Cordes equations. It deserves a serious referee; the construction is novel in this subfield and the core analysis is sound. I would accept it and ask for the Fortin gap and Lemma 11 to be addressed in revision. Yes, I would cite it.","headline":"First ultraweak DPG treatment of Cordes nondivergence-form problems; the continuous theory is sound, but the fully discrete DPG scheme is proven only for piecewise-polynomial A — the DPG-LS variant covers the general case. Worth refereeing, with conditions.","tokens_in":16275,"tokens_out":7582,"would_cite":true,"duration_ms":66805,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N30","65N12"],"pacs":[],"model":"deepseek-v4-flash","headline":"Nondivergence-form PDEs with Cordes coefficients are equivalent to a well-posed ultraweak variational problem, and DPG discretizations of it converge quasi-optimally whenever a bounded Fortin operator exists.","keywords":["ultraweak formulation","DPG method","nondivergence form","Cordes condition","Fortin operator","a posteriori error estimation","adaptive mesh refinement","trace spaces"],"falsifier":"Take a Cordes-coefficient problem with $A$ not piecewise polynomial on any uniformly refined mesh (for instance the radial three-annulus example of Section 6.3) and compare the discrete error ratio $\\|u-u_h\\|_U / \\min_{w_h\\in U_h}\\|u-w_h\\|_U$ along the adaptive sequence; if this ratio grows without bound while the estimator $\\eta$ remains reliable, then the missing Fortin operator is essential rather than a technicality. Conversely, proving the existence of a bounded Fortin operator for every Cordes $A$ would settle the paper's main unresolved limitation in the affirmative.","tokens_in":15180,"feed_emoji":"🧮","tokens_out":10646,"duration_ms":101897,"temperature":0.7,"pith_summary":"This paper proves that a linear second-order PDE in nondivergence form, $Lu = A:D^2u = f$ with zero Dirichlet boundary data, can be rewritten as an ultraweak variational problem when the coefficient matrix $A$ satisfies the Cordes condition. The Hessian $D^2u$ becomes an independent $L^2$-valued unknown $M$, and the continuity of $u$ is carried by a trace variable, so no $H^2$-conforming discrete space is needed. The ultraweak problem is well-posed and equivalent to the original strong problem, and the DPG discretizations inherit quasi-optimal convergence with symmetric positive definite linear systems and local adaptive indicators. Fully discrete quasi-optimality is proven under a Fortin-operator condition; the paper verifies that condition when $A$ is piecewise polynomial of degree $p$, leaving the general Cordes case — including the radial-annulus example of Section 6.3 — as an explicit gap in the theory.","feed_headline":"Nondivergence PDEs fall to a Hessian-as-unknown DPG method","feed_subtitle":"An ultraweak reformulation makes the Hessian an L2 unknown, giving provable convergence when coefficients are piecewise polynomial.","key_machinery":"The mechanism is the broken ultraweak bilinear form $b(u,v) = (u,-\\operatorname{divDiv}Q)_T + (M,Av+Q) + \\langle\\hat u,Q\\rangle_S$ on $U\\times V$, together with the DPG trial-to-test operator $\\Theta$. Elementwise integration by parts moves the second derivatives from $u$ onto the test flux $Q$ and produces the trace pairing $\\langle\\hat u,Q\\rangle_S$; Proposition 3 identifies the trace norm with a supremum over $Q$, and Lemma 11 characterizes the subspace with vanishing normal-normal trace as the annihilator of the trace space. These identities feed the breaking-spaces theorem, which yields well-posedness. For the practical schemes, the Fortin operator $\\Pi_F$ satisfying (10) is the load-bearing device: it preserves the bilinear form on the trial space and is bounded, making the discrete trial-to-test operator $\\Theta_h$ a faithful replacement of $\\Theta$.","core_discovery":"The central claim, stated as Theorem 4 and Proposition 5, is that the strong problem (1) is equivalent to the ultraweak problem (7) and that (7) has a unique solution $u^\\star\\in U$ with $\\|u^\\star\\|_U\\le C\\|f\\|$. Every strong solution $u\\in H^2(\\Omega)\\cap H^1_0(\\Omega)$ maps to the ultraweak solution $(u,D^2u,\\operatorname{tr}^2_T u)$, and every ultraweak solution has a first component in $H^2(\\Omega)\\cap H^1_0(\\Omega)$ whose Hessian is $M$. For the DPG method, the paper proves quasi-optimality of the semi-discrete problem and of the fully discrete problem when a bounded Fortin operator with the Fortin property exists; Theorem 23 constructs such an operator for the lowest-order trial space when $A$ is piecewise polynomial of degree $p$. The paper also proves a posteriori error equivalence, $\\|u-u_h\\|_U^2\\simeq\\eta_{\\mathrm{DPG}}^2$, under the same Fortin assumption.","pith_inferences":["Editorial inference: the open issue is precisely the existence of a Fortin operator for non-piecewise-polynomial Cordes coefficients; if one were constructed, the quasi-optimality and reliability theorems would extend to the radial-annulus example, whose observed adaptive rates in Section 6.3 already match the desired behavior.","Editorial inference: the trace-space template imported from the Kirchhoff–Love plate problem suggests the same ultraweak-plus-trace construction may apply to other linear operators whose highest-order derivative enters linearly, such as fourth-order problems or systems with comparable coefficient structure.","Editorial inference: a direct construction of a Fortin operator for the DPG–least-squares scheme, without assuming $A$ is piecewise polynomial, would also yield one for the DPG method, because Section 2.7 shows the two schemes are equivalent at the operator level; commuting projections on $A$-weighted spaces are a natural starting point."],"forward_implications":["The original nondivergence equation can be approximated with discontinuous piecewise-polynomial spaces for both $u$ and the Hessian $M$, leaving only the trace of the solution to be discretized by H2-type trace elements.","The resulting DPG systems are symmetric positive definite, and the estimator $\\eta_{\\mathrm{DPG}}$ (equivalently $\\eta_{\\mathrm{LS}}$) is equivalent to the $U$-norm error whenever the Fortin assumption holds, so adaptive mesh refinement by local indicators is justified in that regime.","For piecewise-polynomial coefficient matrices aligned with the mesh, the fully discrete DPG and DPG–least-squares methods coincide, so the implementations and their error behavior are interchangeable.","Where the Fortin assumption is not verified, the quasi-optimal bounds (12) and (19) are not in force; Section 6.3 demonstrates that the two fully discrete schemes can then produce different approximations."],"supporting_citations":[{"why":"Supplies the existence, uniqueness, and coercivity of the strong nondivergence problem, which the inf-sup argument of Lemma 2 builds on.","marker":"[22]"},{"why":"Provides the trace operators, the trace-norm identity of Proposition 3, and the rHCT trace space used in the ultraweak formulation.","marker":"[14]"},{"why":"The breaking-spaces theorem that Theorem 4 verifies to prove well-posedness of the ultraweak problem.","marker":"[3]"},{"why":"Source of Lemma 21, the projection $\\Pi_{\\mathrm{divDiv}}$ with the trace and moment properties that define the fully discrete Fortin operator.","marker":"[12]"},{"why":"Gives the Fortin-operator criterion for practical DPG methods used to obtain quasi-optimality of the fully discrete schemes.","marker":"[19]"},{"why":"Introduces the DPG trial-to-test operator that defines the optimal test functions in the semi-discrete method.","marker":"[5]"},{"why":"Supplies the a posteriori equivalence theorem that Theorem 14 adapts to obtain the estimator $\\eta_{\\mathrm{DPG}}$.","marker":"[2]"}],"fun_headline_variants":["Hessian as unknown makes nondivergence PDEs DPG-friendly","Ultraweak DPG: Hessian as L2 unknown for nondivergence PDEs","Provable DPG convergence for nondivergence PDEs with Cordes","Nondivergence PDEs tamed by ultraweak DPG with Hessian unknown"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The fully discrete quasi-optimality bounds and the a posteriori equivalences presuppose a Fortin operator satisfying (10) (or (17)); the paper constructs one only when the coefficient matrix $A$ is piecewise polynomial of degree $p$, so for general Cordes coefficients, including the radial-annulus example of Section 6.3, the convergence theorems as stated do not apply.","fun_headline_variants_meta":{"raw":{"variants":["Hessian as unknown makes nondivergence PDEs DPG-friendly","Ultraweak DPG: Hessian as L2 unknown for nondivergence PDEs","Provable DPG convergence for nondivergence PDEs with Cordes","Nondivergence PDEs tamed by ultraweak DPG with Hessian unknown"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000613,"raw_usage":{"total_tokens":2804,"prompt_tokens":852,"completion_tokens":1952,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":468,"completion_tokens_details":{"reasoning_tokens":1862}},"tokens_in":468,"tokens_out":1952,"duration_ms":14174,"temperature":1.0,"reasoning_tokens":1862,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:25:10.244778+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a Cordes-coefficient problem with $A$ not piecewise polynomial on any uniformly refined mesh (for instance the radial three-annulus example of Section 6.3) and compare the discrete error ratio $\\|u-u_h\\|_U / \\min_{w_h\\in U_h}\\|u-w_h\\|_U$ along the adaptive sequence; if this ratio grows without bound while the estimator $\\eta$ remains reliable, then the missing Fortin operator is essential rather than a technicality. Conversely, proving the existence of a bounded Fortin operator for every Cordes $A$ would settle the paper's main unresolved limitation in the affirmative.","supporting_citations":[{"cited_title":"Smears and E","cited_arxiv_id":null,"evidence_quote":"Supplies the existence, uniqueness, and coercivity of the strong nondivergence problem, which the inf-sup argument of Lemma 2 builds on."},{"cited_title":"Führer, N","cited_arxiv_id":null,"evidence_quote":"Provides the trace operators, the trace-norm identity of Proposition 3, and the rHCT trace space used in the ultraweak formulation."},{"cited_title":"Carstensen, L","cited_arxiv_id":null,"evidence_quote":"The breaking-spaces theorem that Theorem 4 verifies to prove well-posedness of the ultraweak problem."},{"cited_title":"Führer and N","cited_arxiv_id":null,"evidence_quote":"Source of Lemma 21, the projection $\\Pi_{\\mathrm{divDiv}}$ with the trace and moment properties that define the fully discrete Fortin operator."},{"cited_title":"Gopalakrishnan and W","cited_arxiv_id":null,"evidence_quote":"Gives the Fortin-operator criterion for practical DPG methods used to obtain quasi-optimality of the fully discrete schemes."},{"cited_title":"Demkowicz and J","cited_arxiv_id":null,"evidence_quote":"Introduces the DPG trial-to-test operator that defines the optimal test functions in the semi-discrete method."},{"cited_title":"Carstensen, L","cited_arxiv_id":null,"evidence_quote":"Supplies the a posteriori equivalence theorem that Theorem 14 adapts to obtain the estimator $\\eta_{\\mathrm{DPG}}$."}],"review_version":1}