{"id":"c4922036-bb29-467a-a3b7-a630a4193346","arxiv_id":"1908.08732","paper_version":2,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For collocated SBP finite difference operators, discrete Helmholtz Hodge decompositions have an unavoidable remainder caused by grid oscillations, and iterative least-squares projections overcome this in practice.","lead":"This paper proves that a common numerical method for simulating fluids and plasmas cannot exactly split vector fields into swirling and non-swirling parts, because small grid-scale oscillations create an unavoidable remainder. It then provides an iterative fix and demonstrates it on plasma wave analysis.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified","rationale":"The reader's ACCEPT verdict is appropriate. The core proof is a clean dimension argument: for nullspace consistent tensor-product SBP operators, ker grad = span{1}, and the tensor-product structure gives dim ker curl = n + 2 in 3D (and n + 1 in 2D) via rank-nullity plus the fact that coordinate-wise grid oscillations lie in ker grad* and in ker curl. The direct-sum decompositions (14), (19), (52), and (54) are justified by these inclusions and dimension counting. Theorem 5.1's nonvanishing remainder follows because dim(im grad + im curl) <= dim im grad + dim im curl = 3n - 3 < 3n. I checked the weaker spots: Lemma 3.14's commutativity step is valid because the one-dimensional operators and their adjoints commute under tensor products; Corollary 3.16's implicit argument (project out the three oscillation directions, then apply Lemma 3.15) is sound. The nullspace-consistency assumption is indeed the least secure in the proof, but it only controls the explicit form of the extra kernel directions; the strict inequality that powers the impossibility result is robust to larger D-kernels. A minor secondary note: the sequential projection algorithm in Section 5.1 is not shown to produce a residual orthogonal to both im grad and im curl; this would be worth a clarifying remark but does not touch the impossibility theorem.","tokens_in":27005,"tokens_out":30645,"duration_ms":314893,"concrete_test":"Verify the dimension identity behind Theorem 5.1 on a small grid, e.g. N1=N2=N3=2 with the second-order operator of Example 2.2, by explicitly forming the matrices grad and curl and computing the rank of the concatenated matrix [grad, curl]. If the rank is 3n - 3 = 21, the nonvanishing-remainder claim is confirmed in the base case; if the rank is larger, the proof has a hidden flaw.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central negative theorem is internally sound. For nullspace consistent tensor-product SBP operators, the dimension counts in Theorems 3.7, 3.8, 3.18, and 3.19 follow from rank-nullity, commutativity of the discrete derivatives, and the fact that the coordinate-wise grid oscillations lie in the relevant orthogonal complements. Theorem 5.1's nonvanishing remainder is a direct consequence: dim(im grad + im curl) is at most (n-1) + (2n-2) = 3n - 3 < 3n. I checked the key steps (Lemma 3.14, Corollary 3.16, and the direct-sum arguments) and found no gap. The nullspace-consistency premise is the least secure condition in the proof, but even if it failed the strict inequality would typically persist; only the explicit characterization of the extra kernel directions would need revision. Thus no load-bearing concern undermines the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies whether classical vector calculus identities behind the Helmholtz-Hodge decomposition survive for collocated tensor-product finite-difference summation-by-parts (SBP) operators. Under the nullspace-consistency assumption that each one-dimensional derivative satisfies ker D = span{1}, the authors prove that the discrete kernels of curl and div are strictly larger than the images of grad and curl/rot, with the missing dimensions explicitly characterized as grid oscillations, i.e., vectors in ker D*. Consequently, a discrete Poincaré lemma fails, and the discrete Helmholtz-Hodge decomposition u = grad φ + curl v + r has a nonvanishing remainder for certain grid functions. The paper also characterizes fields that are both divergence-free and curl-free as gradients of discrete harmonic functions, proposes a least-squares iterative method (LSQR/LSMR) for computing the decomposition, and supports the theory with numerical experiments showing p+1 convergence of the potentials and vanishing of the remainder under refinement, including an application to MHD wave-mode separation.","tokens_in":116,"tokens_out":30193,"duration_ms":458689,"significance":"The negative result is significant for the SBP community: it shows that classical collocated tensor-product SBP operators cannot exactly mimic the Helmholtz-Hodge decomposition, contrary to what one might expect from their mimetic properties. The obstruction is identified concretely as explicit grid oscillation vectors, and the proofs are elementary finite-dimensional linear algebra. The paper is careful to state its conditional nature, and the numerical experiments with published reproducible code strengthen the contribution. The main limitation is the explicit dependence on nullspace consistency; while the dimension inequality driving the negative result is robust to spurious kernel modes, the exact kernel decompositions would need modification for operators failing Definition 2.3.","major_comments":[{"comment":"Corollary 3.16 is stated without proof, yet it is the step that turns the superset inclusions of Lemmas 3.12 and 3.13 into the exact kernel decompositions of Theorems 3.18 and 3.19. Please add the short argument: by Lemma 3.15, every element of ker curl that is orthogonal to the three vectors (osc1,0,0), (0,osc2,0), (0,0,osc3) lies in im grad; therefore the orthogonal direct sum of im grad with that span has the same dimension as ker curl, namely dim im grad + 3.","section":"Section 3.5, Corollary 3.16"},{"comment":"The proof of Theorem 4.2 asserts that 'the grid oscillations appearing in ker div are not in ker curl and vice versa' implies ker div ∩ ker curl ⊆ im grad ∩ im rot. This implication is not immediate: an element of the intersection could in principle be a sum of a gradient and an oscillator, even if the oscillator itself is not in the other kernel. Please replace this assertion with a direct argument, for example using the orthogonal decompositions of Theorems 3.7–3.19 to show that any component along the oscillator span cannot satisfy the divergence-free condition, so the intersection consists only of gradients (respectively rotations) whose Laplacian vanishes.","section":"Section 4, Theorem 4.2"}],"minor_comments":[{"comment":"The phrase 'the discrete remainder vanishes' could be misread as exact vanishing for finite N, which would contradict Theorem 5.1. Consider rewording to 'the discrete remainder tends to zero under grid refinement'.","section":"Abstract"},{"comment":"In the proof, after stating that the span in (14) is contained in ker grad*, it should also be noted explicitly that the two oscillatory vectors belong to ker curl; this is required for the decomposition and is only implicit.","section":"Section 3.2, proof of Theorem 3.7"},{"comment":"The MATLAB-style pseudo-code for the scaled least-squares solves is helpful, but a short prose explanation of the scaling by sqrtM and why it makes the projections orthogonal with respect to the SBP scalar product would improve readability for readers not familiar with the notation.","section":"Section 5.1"},{"comment":"The observed order 4.6 for the vector potential v with the 2p=6 operator differs from the nominal p+1=4. It would be useful to add one sentence clarifying that this is a single test case and no theoretical claim is made about this particular rate.","section":"Section 6.2"}],"recommendation":"minor_revision","confidential_remarks":"The paper is a solid negative result with a clean and mostly self-contained finite-dimensional proof. The missing proof of Corollary 3.16 and the terse argument in Theorem 4.2 are the only substantive issues; both are local and easily fixed. I have no concerns about novelty, citation practice, or the fit with the journal's scope. The availability of reproducible code is a plus."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid negative-result paper. The central claim is that classical collocated SBP operators cannot satisfy a discrete Poincare lemma or a zero-remainder Helmholtz Hodge decomposition, and the obstruction is characterized explicitly as grid oscillations. I think the claim is right, and the characterization is genuinely new.\n\nWhat is actually new: the kernel decompositions in Theorems 3.7, 3.8, 3.18, and 3.19. They show ker curl = im grad plus explicit oscillatory directions, and similarly for div, under nullspace consistency. The dimension counts are simple but they settle something the SBP/mimetic community has danced around. The paper also proposes a practical workaround: compute least-norm least-squares projections onto im grad and im curl with LSQR/LSMR, and shows numerically that the remainder vanishes under refinement for smooth data. The MHD wave-mode analysis is a nice illustration, though not the main event.\n\nThe proofs are honest finite-dimensional linear algebra; I checked the rank-nullity steps and the direct-sum arguments and found no gaps. The numerical experiments are reproducible, with code on GitHub, and they confirm p+1 convergence of the potentials. Self-citations are to relevant prior work on nullspace consistency and mimetic properties; they are not padding.\n\nSoft spots are minor. The theorems assume nullspace consistency of the one-dimensional D. Some classical operators fail this, and the explicit oscillation characterization would need adjustment there. But the strict dimension inequality in Theorem 5.1 does not depend on the exact form of the oscillations, so the central negative result survives. The vector potential construction is only sketched in Remark 3.17; it would be nice to see it made precise. Remark 5.2 admits that the remainder is not fully characterized, and the observed convergence orders are empirical rather than proven. All of these are stated by the authors; they do not undermine the main theorem.\n\nThis paper is for anyone who uses collocated SBP operators and has wondered whether discrete curl-free fields are gradients. The answer is no, with a precise reason. I would send it to a serious referee — it deserves a proper review rather than a desk reject. My own recommendation would be accept after minor revision, mostly asking for the vector potential sketch to be tightened or explicitly deferred to future work.","headline":"A clean negative theorem: collocated SBP operators fail a discrete Poincare lemma and Helmholtz-Hodge decomposition, with the obstruction characterized exactly as grid oscillations; worth a serious referee.","tokens_in":27645,"tokens_out":2922,"would_cite":true,"duration_ms":30317,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N06","65M06","65N35","65M70","65Z05"],"pacs":[],"model":"deepseek-v4-flash","headline":"For nullspace consistent tensor product SBP operators, the discrete kernel of curl strictly contains the image of grad (and the kernel of div strictly contains the image of curl), so the discrete Helmholtz Hodge decomposition generally…","keywords":["summation by parts","Helmholtz Hodge decomposition","grid oscillations","nullspace consistency","mimetic properties","vector calculus","wave mode analysis"],"falsifier":"Assemble the classical second-order SBP operator of Example 2.2 as a tensor product on a coarse grid such as $N_1 = N_2 = 5$, compute the numerical kernels and ranks of the matrices $\\operatorname{grad}$ and $\\operatorname{curl}$, and check whether $\\dim \\ker \\operatorname{curl} = N_1N_2 + 1$ while $\\dim \\operatorname{im} \\operatorname{grad} = N_1N_2 - 1$; then take $u = (\\mathbf{osc}_{12}, 0)^T$, compute the least-squares projections onto $\\operatorname{im} \\operatorname{grad}$ and $\\operatorname{im} \\operatorname{rot}$, and verify that the remainder $r$ does not vanish. If instead $\\ker \\operatorname{curl} = \\operatorname{im} \\operatorname{grad}$, the paper's central claim is refuted.","tokens_in":26808,"feed_emoji":"📐","tokens_out":14544,"duration_ms":122813,"temperature":0.7,"pith_summary":"This paper asks whether the classical representation theorems of vector calculus survive discretization by finite difference summation by parts (SBP) operators: every curl-free vector field is a gradient, every divergence-free field is a curl, and every field splits orthogonally into a gradient plus a curl (the Helmholtz Hodge decomposition). The answer is negative for the classical collocated operators: under nullspace consistency, the discrete kernel of curl is strictly larger than the image of grad, and the kernel of div is strictly larger than the image of curl, in both two and three dimensions. The extra dimensions are spanned by explicit grid oscillation vectors, so some irrotational or solenoidal grid functions have no scalar or vector potential, and the discrete Helmholtz Hodge decomposition generally leaves a nonzero remainder. The paper then shows the obstruction does not prevent the decomposition from being useful: least-squares projections onto the gradient and curl images converge, the remainder shrinks under grid refinement for smooth data, and the machinery separates magnetohydrodynamic wave modes in practice.","feed_headline":"Grid oscillations break discrete Helmholtz Hodge decompositions","feed_subtitle":"On standard SBP grids, curl-free discrete fields need not be gradients; explicit oscillatory modes fill the gap.","key_machinery":"The load-bearing object is the nullspace consistent summation by parts (SBP) derivative operator: a discrete derivative $D$ together with a symmetric positive definite mass matrix $M$ and boundary operator $E$ satisfying $MD + D^T M = E$, with the additional requirement $\\ker D = \\operatorname{span}\\{\\mathbb{1}\\}$. For such an operator the adjoint $D^* = M^{-1}D^T M$ has a one-dimensional kernel spanned by a grid oscillation vector $\\mathbf{osc}$, which alternates between positive and negative values across the grid. In tensor product dimensions, these oscillation vectors placed in single coordinate components generate exactly the gap between $\\operatorname{im} \\operatorname{grad}$ and $\\ker \\operatorname{curl}$ (and between $\\operatorname{im} \\operatorname{curl}$ and $\\ker \\operatorname{div}$). Two tools carry the proofs: dimension counts relating $\\ker D^*$ to $\\ker D$, and explicit discrete inverse operators $D^{-1}$ that integrate along coordinate lines to build scalar potentials for fields orthogonal to the oscillations. The same machinery produces the filter operator $F$ that projects grid oscillations out of a field, and the least-squares projection setup used for the numerical decompositions.","core_discovery":"The central discovery is a dimension mismatch with explicit generators. For nullspace consistent tensor product SBP operators, $\\dim \\operatorname{im} \\operatorname{grad} = N_1N_2 - 1 < N_1N_2 + 1 = \\dim \\ker \\operatorname{curl}$ in two dimensions, with $\\ker \\operatorname{curl} = \\operatorname{im} \\operatorname{grad} \\oplus \\operatorname{span}\\{(\\mathbf{osc}_1, 0)^T, (0, \\mathbf{osc}_2)^T\\}$; in three dimensions $\\dim \\ker \\operatorname{curl} = N_1N_2N_3 + 2$ versus $\\dim \\operatorname{im} \\operatorname{grad} = N_1N_2N_3 - 1$, and the corresponding statements hold for divergence versus curl (Theorems 3.7, 3.8, 3.18, 3.19). Here $\\mathbf{osc}_i$ is the one-dimensional grid oscillation vector spanning the kernel of the adjoint derivative $D_i^*$, a vector that alternates in sign across the grid. Because the kernels are strictly larger than the images, a discrete Poincaré lemma fails: there are grid functions $u$ that cannot be written as $\\operatorname{grad}\\phi + \\operatorname{curl} v$, so the discrete Helmholtz Hodge decomposition $u = \\operatorname{grad}\\phi + \\operatorname{curl} v + r$ has a nonzero remainder $r$ lying in the orthogonal complement of both image spaces (Theorem 5.1). The paper's positive claim is that the obstruction is numerically benign: computing the decomposition as mass-matrix-scaled least-norm least-squares projections (LSQR in 2D, LSMR in 3D), the potentials and components converge with the expected order of accuracy, the remainder vanishes under refinement for smooth data, and the method separates MHD wave modes when the projection order is matched to the dominant amplitude.","pith_inferences":["The same kernel mismatch should appear in any collocated discretization whose derivative adjoint has a nontrivial kernel, including global spectral collocation at Lobatto–Legendre nodes, where the oscillation vector is the highest Legendre mode; the paper notes this class satisfies its assumptions, so the negative result likely transfers verbatim.","A practical adaptive rule suggested by the MHD experiments — estimate the amplitudes of the competing wave families and project first onto the dominant one — could be formalized and tested, since the paper only reports the qualitative direction of the effect, not a threshold.","A rigorous convergence rate for the remainder $\\|r\\|_M$ under grid refinement is a natural open problem; the paper reports experimental orders of accuracy ($p+1$, with one component at $4.6$ for the sixth-order operator) but proves no error estimate.","Because staggered-grid and mimetic operators do support exact discrete decompositions, as the paper notes, moving to staggered SBP schemes is a concrete, testable remedy for applications that need a discrete Poincaré lemma to hold exactly."],"forward_implications":["Irrotational discrete vector fields need not be gradients: for nullspace consistent tensor product SBP operators, $\\ker \\operatorname{curl}$ strictly contains $\\operatorname{im} \\operatorname{grad}$ in two and three dimensions, so a discrete Poincaré lemma fails.","The discrete Helmholtz Hodge decomposition $u = \\operatorname{grad}\\phi + \\operatorname{curl} v + r$ generally has a nonzero remainder $r$, orthogonal to both image spaces; this is a structural property of collocated SBP operators, not a numerical defect.","For grid functions orthogonal to the one-dimensional grid oscillations, the classical potential theorems hold, and the paper's filter operator $F$ restores that situation by explicitly removing the oscillations.","Because $\\operatorname{im} \\operatorname{grad}$ and $\\operatorname{im} \\operatorname{curl}$ are not orthogonal, the two projection orders yield different decompositions, and the MHD tests show the correct order depends on which wave amplitude dominates.","In numerical experiments the decomposition converges: potentials and components reach order $p+1$ accuracy for operators of interior order $2p$, the remainder vanishes under refinement for smooth data, and MHD wave-mode separation succeeds when the projection order matches the dominant amplitude."],"supporting_citations":[{"why":"supplies the continuous Helmholtz Hodge decompositions and potential theorems whose discrete analogues the paper proves fail","marker":"[18]"},{"why":"gives the continuous equivalence between divergence/curl-free fields and gradients of harmonic functions used as Theorem 4.1","marker":"[51]"},{"why":"originates finite difference summation by parts operators and provides the defining SBP property (2.1)","marker":"[29]"},{"why":"founds the SBP framework for d/dx that the tensor product constructions build on","marker":"[55]"},{"why":"provides the classical second-through-eighth order SBP operators used in all numerical experiments","marker":"[36]"},{"why":"introduces the nullspace consistency property that the paper's theorems assume","marker":"[32]"},{"why":"supplies the LSQR iterative least-squares solver used for the 2D projections","marker":"[41]"},{"why":"supplies the LSMR iterative least-squares solver used for the 3D projections","marker":"[15]"},{"why":"provides the contrasting mimetic finite difference decomposition theorems that hold for staggered operators, not collocated ones","marker":"[26]"}],"fun_headline_variants":["Grid oscillations create spurious Hodge remainder","Discrete Hodge decomposition: missing potentials, but iterative fix","Helmholtz Hodge on SBP grids: gap closed by iterations","SBP discrete calculus: Hodge fails, LSQR/LSMR succeed","Oscillatory modes break Hodge; projection method repairs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"All of the theorems assume nullspace consistency: each one-dimensional SBP derivative operator $D$ must have kernel exactly $\\operatorname{span}\\{\\mathbb{1}\\}$, so that spurious kernel modes are absent; if an operator has extra kernel modes, the dimension counts and the explicit spanning sets no longer follow.","fun_headline_variants_meta":{"raw":{"variants":["Grid oscillations create spurious Hodge remainder","Discrete Hodge decomposition: missing potentials, but iterative fix","Helmholtz Hodge on SBP grids: gap closed by iterations","SBP discrete calculus: Hodge fails, LSQR/LSMR succeed","Oscillatory modes break Hodge; projection method repairs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000536,"raw_usage":{"total_tokens":2660,"prompt_tokens":1118,"completion_tokens":1542,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":734,"completion_tokens_details":{"reasoning_tokens":1455}},"tokens_in":734,"tokens_out":1542,"duration_ms":12253,"temperature":1.0,"reasoning_tokens":1455,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:31:33.545914+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Assemble the classical second-order SBP operator of Example 2.2 as a tensor product on a coarse grid such as $N_1 = N_2 = 5$, compute the numerical kernels and ranks of the matrices $\\operatorname{grad}$ and $\\operatorname{curl}$, and check whether $\\dim \\ker \\operatorname{curl} = N_1N_2 + 1$ while $\\dim \\operatorname{im} \\operatorname{grad} = N_1N_2 - 1$; then take $u = (\\mathbf{osc}_{12}, 0)^T$, compute the least-squares projections onto $\\operatorname{im} \\operatorname{grad}$ and $\\operatorname{im} \\operatorname{rot}$, and verify that the remainder $r$ does not vanish. If instead $\\ker \\operatorname{curl} = \\operatorname{im} \\operatorname{grad}$, the paper's central claim is refuted.","supporting_citations":[{"cited_title":"Schweizer","cited_arxiv_id":null,"evidence_quote":"gives the continuous equivalence between divergence/curl-free fields and gradients of harmonic functions used as Theorem 4.1"},{"cited_title":"Kreiss and G","cited_arxiv_id":null,"evidence_quote":"originates finite difference summation by parts operators and provides the defining SBP property (2.1)"},{"cited_title":"Linders, J","cited_arxiv_id":null,"evidence_quote":"introduces the nullspace consistency property that the paper's theorems assume"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the LSQR iterative least-squares solver used for the 2D projections"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the LSMR iterative least-squares solver used for the 3D projections"}],"review_version":1}