{"id":"98273c50-a5bf-4f5a-8513-7539f65ef15d","arxiv_id":"2411.14561","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Three preconditioners for H(div) interior penalty DG are constructed and analyzed, with provably h-, p-, and eta-robust conditioning on affine and Cartesian meshes.","lead":"This paper designs and analyzes three preconditioners for interior penalty discontinuous Galerkin discretizations posed in the H(div) finite element space, the velocity space used in exactly divergence-free Stokes solvers. On affine or Cartesian meshes, two of the preconditioners provably converge with iteration counts independent of mesh size, polynomial degree, and penalty parameter.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 9's cross-term counting treats the global coarse space V0 as one additional neighbor, but V0 overlaps every vertex patch, so the Young's inequality step as written yields a K-dependent constant.","rationale":"The reader's weakest_assumption correctly identifies the coarse-space counting in Lemma 9 as the softest spot in the proof of Theorem 1. I agree with this diagnosis: the text's 'plus one for the coarse space' is only valid for local-vs-local cross terms, and the coarse space's interaction with all vertex patches is not addressed. However, a standard grouped Young's inequality argument repairs the lemma without changing the constants' dependence on h, p, or eta. Because the gap is a proof omission rather than a mathematical contradiction, and because the numerical results (Tables 1-4) support the claimed behavior, the appropriate verdict remains CONDITIONAL: the central claims are credible but the manuscript needs the missing argument to be fully established. I therefore keep the reader's verdict unchanged. A secondary concern about the affine-mesh jump estimate in Lemma 10 is also present, but the coarse-space counting is the most directly load-bearing for the h-independence claim and is the one explicitly flagged by the reader.","tokens_in":27804,"tokens_out":17472,"duration_ms":166910,"concrete_test":"Rewrite the proof of Lemma 9 using the grouped decomposition: for any vh = v0 + v_rest with v_rest = sum_{i>=1} vi, apply 2a(v0,v_rest) <= (1/2)a(v0,v0) + 2a(v_rest,v_rest), then bound a(v_rest,v_rest) <= c sum_{i>=1} a(vi,vi) using the bounded valence of the mesh. Verify that the resulting constant c is independent of the number of vertices K and of h. If this verification succeeds, Theorem 1's h-independence is restored; if it fails, the theorem is unproved for general meshes as written.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Lemma 9 (Section 3.1) proves the lower spectral bound for T^{-1} by bounding a(vh,vh) above by a constant times the sum of the subspace energies for an arbitrary decomposition. The textual counting says the number of nonzero cross terms a(vi,vj) is bounded by the mesh valence 'plus one for the coarse space V0.' This is only valid for pairs of vertex-patch spaces (i,j >= 1). For i = 0, the global coarse space V0 has non-negligible a-inner products with every one of the K vertex-patch spaces Vi, so the number of cross terms involving V0 grows with the number of vertices. Applying Cauchy-Schwarz and Young's inequality separately to each of those K terms would introduce a factor K in the constant, breaking the claimed h-independence of Theorem 1. The gap is likely repairable: group all local components into v_rest = sum_{i>=1} vi, bound the single cross term a(v0, v_rest) by (1/2)a(v0,v0) + (1/2)a(v_rest,v_rest) via Young's inequality, and then use the finite-overlap/valence argument only among the local spaces to control a(v_rest,v_rest). This yields a constant independent of K. However, this argument is absent from the manuscript, so the proof as written does not establish the lower bound on general meshes.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops and analyzes three preconditioners for the interior penalty discontinuous Galerkin discretization of a vector Poisson problem posed in the H(div)-conforming Raviart-Thomas space: an additive subspace correction preconditioner using vertex patches and a lowest-order H1-conforming coarse space, a fictitious space preconditioner using the degree-p discontinuous Galerkin space, and an auxiliary space preconditioner using the degree-(p-1) discontinuous Galerkin space with a block Jacobi smoother. The main theoretical claims are that the subspace correction preconditioner has a condition number independent of h, and also independent of p and eta on affine meshes (Theorem 1); the fictitious space preconditioner is h-robust in general and p/eta-robust on Cartesian meshes (Theorem 2); and the auxiliary space preconditioner is h-robust on general meshes (Theorem 3). Section 4 gives explicit Gauss-Lobatto interpolation stability and error estimates with sharp constants. Numerical experiments on Cartesian, affine unstructured, and non-affine skewed meshes, together with block-diagonal preconditioning of an exactly divergence-free Stokes discretization, support the predicted robustness and show mild p-dependence in the cases without full theoretical guarantees.","tokens_in":28091,"tokens_out":15773,"duration_ms":163285,"significance":"If the theoretical claims are fully established, the paper makes a useful contribution to scalable solvers for H(div)-conforming DG methods, which are important for pressure-robust, exactly divergence-free discretizations of incompressible flow. The matrix-free implementation of the fictitious space preconditioner using low-order-refined spectral equivalence is practically relevant, and the explicit Gauss-Lobatto interpolation estimates in Section 4 are a useful technical contribution. The numerical study is reasonably broad, covering parameter sweeps in h, p, and eta on three mesh families as well as Stokes/MINRES. The main weaknesses are several proof gaps in the general-mesh parts of Theorems 1-3; these concern the coarse-space interaction in Lemma 9, the H1-stability step applied to a non-H1 function in Lemma 10, and the interpolation estimates for non-polynomial functions in the proof of Theorem 3. The affine and Cartesian cases are better supported by the text.","major_comments":[{"comment":"Lemma 9 (Section 3.1, Eq. (14)): the lower spectral bound is proved by counting non-negligible cross terms a(vi,vj) and claiming that the count is bounded by the mesh valence plus one for the coarse space V0. This is only valid for pairs of vertex-patch spaces. The global coarse space V0 has non-negligible a-inner products with every vertex-patch space Vi, so a termwise application of Cauchy-Schwarz and Young introduces a factor depending on the number of vertices K. The resulting lower bound is therefore not h-independent as written. The gap is repairable by grouping all local components into v_rest = sum_{i>=1} vi, applying Young once to a(v0, v_rest), and then using finite overlap only among the local spaces, but this argument is absent from the manuscript. As written, the proof does not establish the h-independent lower bound that Theorem 1 requires.","section":"Section 3.1, Lemma 9"},{"comment":"Lemma 10 (Section 3.1): in the general non-affine case, the proof bounds the coarse term by asserting H1 stability of the nodal interpolant IV0 applied to Qh(vh). However, Section 2.2 explicitly states that Qh(vh) belongs to [H1]^d only when all mesh elements are affine; on non-affine meshes this is not generally true. Thus the argument that |||v0|||^2 is controlled by |||vh|||^2 is not justified for the general meshes covered by the h-independence part of Theorem 1. The affine case is fine, but the general-mesh statement needs a separate argument, for example one that estimates the interpolation of the possibly discontinuous averaged field directly through its jumps.","section":"Section 3.1, Lemma 10"},{"comment":"The proof of the c0 bound in Theorem 3 sets v0 = I_W0(v) and invokes Lemma 6 and a scaling argument to obtain both stability of I_W0 on v and the estimate ||v - v0||_0^2 ≲ (h^2/p^2)||∇_h v||_0^2. On non-affine elements, a function v in Vh is not a polynomial; it is the Piola image of a polynomial under a spatially varying Jacobian, and the interpolation operator I_W0 into Q_{p-1}^d is not Piola-covariant. The cited one-dimensional polynomial interpolation results therefore do not apply directly. Since Theorem 3 only claims h-independence, constants depending on p and eta would be acceptable, but the estimate must be proved. As written, the h-independence of c0 on general meshes is not established.","section":"Section 3.3, proof of Theorem 3"}],"minor_comments":[{"comment":"In Eq. (16) and the surrounding text, the symbol vj appears where vh is clearly intended; this makes the proof harder to follow.","section":"Section 3.1, Lemma 10"},{"comment":"The right-inverse S is introduced only in passing, and the phrase Svj appears to be a typo for Svh. Please define S explicitly and state how the scaling argument is uniform in h on non-affine elements.","section":"Section 3.2, proof of Theorem 2"},{"comment":"The statement says For all v in V(kappa) but the displayed inequalities use u; please align the notation.","section":"Lemma 3"},{"comment":"In the final display of the proof, the interpolation operator is denoted J_{p-1} instead of I_{p-1}; this is a typo.","section":"Section 4, proof of Lemma 15"},{"comment":"The caption contains the duplicated phrase with with non-affine; it should read with non-affine elements.","section":"Section 5.3, Figure 1 caption"},{"comment":"The phrase On meshes with non-affine meshes is redundant; change it to On non-affine meshes.","section":"Remark 3"}],"recommendation":"major_revision","confidential_remarks":"The numerical work appears solid and the topic fits the journal. My recommendation is based on proof gaps rather than on the numerics. The use of the author's previous results, including [34,35,36], is appropriate and does not create circularity. I do not see a problem with the intermediate eta p^4 factor in the proof of Theorem 2: it cancels with the h/p^2 estimate from Lemma 4. The main issues to be addressed are the coarse-space counting in Lemma 9, the H1-stability step in Lemma 10 for non-affine meshes, and the non-affine interpolation estimates in the proof of Theorem 3."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a good paper that does something genuinely new—it carries subspace correction, fictitious space, and auxiliary space preconditioning over to H(div)-conforming interior penalty DG, with Piola-mapped Raviart–Thomas elements and explicit Gauss–Lobatto constants. The numerical study is thorough, covers structured and unstructured meshes, and includes the Stokes application. The central claims are believable and, I think, correct. But the proofs have real gaps, and the most important one is in Lemma 9.\n\nThe Lemma 9 issue is exactly as your stress-test note says. The proof counts cross terms a(vi,vj) as bounded by mesh valence \"plus one for the coarse space V0.\" That works for pairs of vertex patches, but V0 overlaps every one of the K local spaces, so the number of nonzero terms involving V0 grows with K. As written, the Young's inequality step would introduce a K-dependent constant and break the h-independence claim. The repair is straightforward—group all local components into v_rest and apply Young once to a(v0,v_rest), then use the finite-overlap argument only among local spaces—but the argument as printed does not establish the lower bound on general meshes. This is a load-bearing proof gap, not a cosmetic one.\n\nI'd also flag Lemma 10. The proof invokes H1 stability of the vertex interpolant IV0 applied to Qh(vh). On non-affine meshes Qh(vh) is not necessarily in H1, so the H1 stability statement needs qualification. You can probably fix it with a broken-norm or jump-aware bound, but again, it is not written.\n\nOn Theorem 2, I disagree with the reader's flagged \"eta p^4\" issue. That factor appears in the intermediate inequality but cancels against the h/p^2 bound from Lemma 4, leaving exactly the jump term that is part of the DG norm. So I do not see a real problem there.\n\nWhat the paper does well: it clearly separates affine from non-affine cases, gives explicit constants where it can, and is honest about where theory stops (Remarks 3 and 5). The numerical results are strong and support the claims of h-robustness on general meshes and mild p-dependence. The paper is well-written and will be useful to people building solvers for exactly divergence-free Stokes discretizations.\n\nBottom line: this deserves a serious referee. The gaps are repairable, but they should be fixed before publication—especially Lemma 9. I would send it out, with a request that the author address the coarse-space counting and the H1-stability step. A careful reader could get the proofs into good shape, and the paper would be a solid contribution to the H(div) preconditioning literature.","headline":"A solid, useful preconditioning paper for H(div)-conforming IPDG with convincing numerics and a few repairable proof gaps, most notably the coarse-space counting in Lemma 9.","tokens_in":28586,"tokens_out":2778,"would_cite":true,"duration_ms":28308,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N30","65N22","65F08","65N55"],"pacs":[],"model":"deepseek-v4-flash","headline":"Vertex-patch, fictitious-space, and auxiliary-space preconditioners for H(div) DG discretizations yield condition numbers independent of mesh size and, on affine meshes, of polynomial degree and penalty parameter.","keywords":["H(div) finite elements","interior penalty discontinuous Galerkin","subspace correction preconditioner","fictitious space preconditioner","auxiliary space preconditioner","vertex patch","Gauss-Lobatto interpolation","Stokes saddle-point preconditioning"],"falsifier":"Run the subspace-correction preconditioner on a family of meshes formed by subdividing a disk into $N$ sectors meeting at a single central vertex, so that the central vertex's valence grows with $N$, and plot the condition number of $BA$ against $N$; a condition number that grows with $N$ (or with the total vertex count on any fixed-valence refinement) would falsify the claimed $h$-independence.","tokens_in":27608,"feed_emoji":"🧮","tokens_out":16172,"duration_ms":121653,"temperature":0.7,"pith_summary":"This paper builds and analyzes preconditioners for the interior penalty discontinuous Galerkin discretization of the vector Poisson problem posed in $H(\\mathrm{div})$, the finite element space of functions with continuous normal components. Such discretizations are the velocity block in exactly divergence-free, pressure-robust Stokes solvers, so a good preconditioner there unlocks scalable incompressible-flow simulation. The paper claims three preconditioners—a subspace correction method using vertex patches and a lowest-order $H^1$-conforming coarse space, a fictitious space method using the degree-$p$ discontinuous Galerkin space, and an auxiliary space method using the degree-$(p-1)$ DG space with a block Jacobi smoother—produce preconditioned systems whose condition numbers are independent of mesh size $h$; on affine (parallelogram) meshes, the subspace and fictitious space methods are additionally independent of polynomial degree $p$ and penalty parameter $\\eta$. Numerical experiments on structured and unstructured meshes show iteration counts that grow only mildly with $p$ even when the theory does not guarantee $p$-independence, and the methods drive uniform MINRES convergence when used as the velocity block of a block-diagonal Stokes preconditioner.","feed_headline":"H(div) DG solvers get condition numbers independent of h and p","feed_subtitle":"Subspace, fictitious, and auxiliary space solvers keep CG iteration counts flat on structured and unstructured meshes.","key_machinery":"The load-bearing object is the subspace decomposition $V_h = V_0 + \\sum_i V_i$, where $V_0$ is the lowest-order $H^1$-conforming space and $V_i$ are vertex-patch spaces of Raviart-Thomas functions supported on the union of elements around each vertex. The argument hinges on two inequalities: a finite-overlap bound that limits the cross terms among the $V_i$ in the $a(\\cdot,\\cdot)$ inner product, and a stable-decomposition estimate that builds an explicit split of any $v_h$ into a coarse part (via Oswald averaging and vertex interpolation) and local parts (via Gauss-Lobatto nodal interpolation of vertex-patch products), with all constants controlled by the explicit Gauss-Lobatto interpolation stability lemmas in Section 4. The fictitious space and auxiliary space methods reuse the same interpolation and averaging machinery through the transfer operator $R = \\hat{Q} I_{\\hat{V}_h}$, where $I_{\\hat{V}_h}$ is nodal interpolation into the broken $H(\\mathrm{div})$ space and $\\hat{Q}$ is a diagonal averaging operator, reducing the $H(\\mathrm{div})$ problem to a standard DG problem whose inverse can be approximated by spectrally equivalent low-order refined preconditioners.","core_discovery":"The paper's central claim is Theorem 1: for the subspace correction preconditioner $B$, the condition number of the preconditioned operator $BA$ is bounded independently of the mesh size $h$, and when all mesh elements are affinely mapped, independently of the polynomial degree $p$ and the penalty parameter $\\eta$. Theorems 2 and 3 extend the $h$-independence to a fictitious space preconditioner built from the degree-$p$ discontinuous Galerkin space and to an auxiliary space preconditioner built from the degree-$(p-1)$ DG space with a block Jacobi smoother, with $p$- and $\\eta$-independence shown on Cartesian (or affinely transformed Cartesian) meshes. The analysis goes through a stable subspace decomposition into vertex patches plus a coarse space, with the stability constants controlled by explicit Gauss-Lobatto interpolation estimates, and through a transfer operator between the $H(\\mathrm{div})$ space and the piecewise polynomial DG space. On general non-affine meshes the paper proves only $h$-independence, but reports numerical evidence that $p$-dependence stays mild.","pith_inferences":["A strict reading of Lemma 9 suggests the proof's overlap count treats the coarse space $V_0$ as a single additional overlap, but $V_0$ interacts with every vertex patch; establishing the full $h$-independence on general meshes would need a separate estimate for coarse-to-patch cross terms.","The explicit Gauss-Lobatto constants in Section 4 give the unproved bound $\\kappa \\lesssim \\eta p^2$ for the subspace preconditioner on non-affine meshes, so explaining the observed mild numerical growth would require either sharper trace estimates or a different vertex-patch decomposition.","The fictitious space construction naturally suggests a full $p$-multigrid cycle by recursively applying the degree-$(p-1)$ auxiliary space idea, an extension the paper does not pursue.","Because the low-order refined operator used in the fictitious space preconditioner is an M-matrix written as a weighted graph Laplacian, classical algebraic multigrid theory could be invoked to prove uniform convergence of the approximate solver, not just the ideal inverse, on general meshes."],"forward_implications":["If the $h$-independence holds as stated, the subspace correction preconditioner becomes a drop-in replacement for the velocity-block inverse in block-diagonal Stokes preconditioners, making the full saddle-point solve uniform in mesh size without global iterative tuning.","The fictitious space preconditioner, being expressible through Kronecker-product sum factorization and a low-order refined DG operator, can be implemented matrix-free with $O(p^{d+1})$ work and $O(p^d)$ memory per element, so high-order $H(\\mathrm{div})$ discretizations become practical at very high $p$.","The auxiliary space preconditioner, which avoids the expensive vertex-patch solves, offers a cheaper route to $p$-robust convergence on unstructured meshes, and the paper's numerical results indicate it is the fastest of the three at high order on non-affine meshes.","On affine meshes, independence from the penalty parameter $\\eta$ means users are free to choose $\\eta$ for stability or accuracy without worrying about solver degradation, which the paper notes is especially relevant for the $\\eta$-sensitive inf-sup constant in Stokes."],"supporting_citations":[{"why":"Supplies the subspace-correction duality identity (Lemma 8) that translates the stable decomposition into a condition-number bound.","marker":"[48]"},{"why":"Establishes the $H^1$-stability of Gauss-Lobatto interpolation on vertex patches that underlies the stable decomposition for the subspace preconditioner.","marker":"[32]"},{"why":"Provides the analogous p-version domain-decomposition stability results referenced alongside [32].","marker":"[33]"},{"why":"Gives the fictitious-space lemma (Lemma 11) that yields the condition-number bound of the fictitious-space preconditioner from the two constants $c_R$ and $c_S$.","marker":"[28]"},{"why":"Supplies the auxiliary-space convergence theorem (Lemma 14) used to bound the auxiliary space preconditioner.","marker":"[45]"},{"why":"Provides the trace and inverse trace inequalities plus the Oswald-averaging estimate that control the remainder terms in the stable decomposition.","marker":"[14]"},{"why":"Supplies the low-order-refined spectral equivalence used to replace the DG inverse in the fictitious space preconditioner with an optimal-cost M-matrix problem.","marker":"[36]"},{"why":"Introduces the exactly divergence-free $H(\\mathrm{div})$ DG discretization of Stokes that the preconditioners target as the velocity-block solver.","marker":"[15]"}],"fun_headline_variants":["H(div) DG preconditioners: uniform conditioning in h and p","Mesh and degree robust H(div) DG preconditioners","Condition number independent of h and p for H(div) DG","Three preconditioners make H(div) DG well-conditioned"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The $h$-independence bound depends on the subspace overlap counting staying bounded by a constant, but the coarse space $V_0$ overlaps with every vertex patch, and the paper provides no separate estimate for those coarse-to-patch cross terms.","fun_headline_variants_meta":{"raw":{"variants":["H(div) DG preconditioners: uniform conditioning in h and p","Mesh and degree robust H(div) DG preconditioners","Condition number independent of h and p for H(div) DG","Three preconditioners make H(div) DG well-conditioned"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000832,"raw_usage":{"total_tokens":3665,"prompt_tokens":1014,"completion_tokens":2651,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":630,"completion_tokens_details":{"reasoning_tokens":2587}},"tokens_in":630,"tokens_out":2651,"duration_ms":18016,"temperature":1.0,"reasoning_tokens":2587,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:10:19.120039+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the subspace-correction preconditioner on a family of meshes formed by subdividing a disk into $N$ sectors meeting at a single central vertex, so that the central vertex's valence grows with $N$, and plot the condition number of $BA$ against $N$; a condition number that grows with $N$ (or with the total vertex count on any fixed-valence refinement) would falsify the claimed $h$-independence.","supporting_citations":[{"cited_title":"Domain decomposition algorithms for the p-version finite element method for elliptic problems","cited_arxiv_id":null,"evidence_quote":"Provides the analogous p-version domain-decomposition stability results referenced alongside [32]."},{"cited_title":"Decomposition and fictitious domains meth ods for elliptic boundary value prob- lems","cited_arxiv_id":null,"evidence_quote":"Gives the fictitious-space lemma (Lemma 11) that yields the condition-number bound of the fictitious-space preconditioner from the two constants $c_R$ and $c_S$."},{"cited_title":"Low-order precon ditioning for the high-order finite element de Rham complex","cited_arxiv_id":null,"evidence_quote":"Supplies the low-order-refined spectral equivalence used to replace the DG inverse in the fictitious space preconditioner with an optimal-cost M-matrix problem."},{"cited_title":"A note on disc ontinuous Galerkin divergence-free solu- tions of the Navier–Stokes equations","cited_arxiv_id":null,"evidence_quote":"Introduces the exactly divergence-free $H(\\mathrm{div})$ DG discretization of Stokes that the preconditioners target as the velocity-block solver."}],"review_version":1}