{"id":"656ef0e9-f3cf-4590-be26-579a51d33527","arxiv_id":"1908.04498","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Auxiliary space preconditioners of the form ∇*B_div∇ are shown to be uniformly spectrally equivalent to the negative-order fractional Laplacian; the required fractional H(div) preconditioner is built by additive multigrid but depends on an unproven two-level estimate.","lead":"This paper builds fast numerical solvers (preconditioners) for equations involving the fractional Laplacian with negative exponent, a tool used in coupled physics problems with interface constraints. The author shows one can construct such solvers by combining gradient operators with preconditioners for a related fractional H(div) operator, and tests the approach on model problems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The additive-multigrid construction of B_{div,h}^{1-s} rests on the two-level estimates in Lemma 4.2, equation (4.20), which the paper explicitly leaves unproven; the numerical experiments with fixed J=4 do not close that gap.","rationale":"The reader's verdict identifies the unproven two-level estimates (4.20) as the weakest point, and my reading confirms that this is the single most load-bearing concern. The auxiliary-space theorem (Corollary 3.1) is proven rigorously and is well supported by the numerical condition numbers in Table 2, which match β^{-2(1+s)} to three decimals. However, the construction of the fractional H(div) preconditioner, which is an essential input to that theorem, depends on Lemma 4.2's hypotheses, and those hypotheses are explicitly unproven in Remark 3. The numerical experiments in Section 5 are encouraging but not dispositive: they keep J=4 fixed, so they establish h-independence at fixed multigrid depth but not the level-independence needed for the two-level estimate. Since the paper is transparent about the gap, the appropriate verdict remains CONDITIONAL rather than REJECT, and my stress-test does not change the reader's assessment.","tokens_in":21155,"tokens_out":4129,"duration_ms":43416,"concrete_test":"On a sequence of uniformly refined meshes with levels k=2,...,8 using lowest-order Raviart-Thomas elements, generate many τ in (I−P^s_{k,k−1})V_k, compute the discrete Helmholtz decomposition τ=∇_k u+curl q, and evaluate R1 = ‖∇_k u‖² / (h_{k−1}^{2s}(Λ_k^s τ,τ)) and R2 = ‖q‖ / (h_{k−1}‖τ‖) for, say, s=0.25, 0.5, 0.75. If either ratio grows with k or exceeds a modest constant, then Assumption (4.20) fails and Lemma 4.2 collapses; if both remain uniformly bounded, the missing step is at least numerically credible, though an analytical proof would still be required.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The Section 3 auxiliary-space argument is sound conditional on having an operator B_{div,h}^{1-s} satisfying the two-sided spectral bound (3.24). The load-bearing step is Section 4's claim to supply such an operator via the additive multigrid operator (4.6). The lower bound in Theorem 4.1 requires Assumption (4.8), and the only route to (4.8) offered is Lemma 4.2. Lemma 4.2 is itself conditional on the two-level estimates in (4.20): for every τ in (I−P^s_{k,k−1})V_k with Helmholtz decomposition τ=∇_k u+curl q, one needs ‖∇_k u‖² ≤ c h_{k−1}^{2s}(Λ_k^s τ,τ) and ‖q‖ ≤ c h_{k−1}‖τ‖. Remark 3 states explicitly that verifying these estimates is left as future work, so the paper does not prove that B_{div,h}^{1-s} satisfies (3.24). This is an omitted proof rather than an internal contradiction, but it is essential: without (4.20), Theorem 4.1 does not deliver spectral equivalence, and the abstract's claim that such operators are 'designed' is not fully supported. The numerical tables are consistent with the theory, but they use four levels J=4 while increasing N, so they do not exercise the level-independence of the two-level estimate and cannot substitute for the missing proof.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an auxiliary-space preconditioner for discretizations of the fractional Laplacian of negative order. Section 3 develops a clean argument, both continuous and discrete, showing that B_s^h = ∇_h^* B_{div,h}^{1-s} ∇_h is spectrally equivalent to A_h^{-s} provided B_{div,h}^{1-s} satisfies the two-sided bound (3.24). Section 4 attempts to supply such a B_{div,h}^{1-s} as the additive multigrid operator (4.6). Theorem 4.1 gives sufficient conditions on the smoothers, and Lemma 4.2 verifies the main condition under two-level error estimates (4.20) that Remark 3 explicitly leaves unproven. Numerical experiments with a fixed number of levels (J=4) report bounded iteration counts and condition numbers.","tokens_in":21608,"tokens_out":7158,"duration_ms":75640,"significance":"The Section 3 transfer argument is elegant, correctly uses interpolation and Jensen's operator inequality, and is of independent interest. If the missing estimates were supplied, the paper would give a useful preconditioning strategy for the fractional interface problems that motivate it. The author is clearly aware of the main theoretical gap and states it honestly. However, as submitted the central construction of B_{div,h}^{1-s} is conditional, and the numerical experiments, while consistent with the theory, do not test the level-dependence that the theory still leaves open.","major_comments":[{"comment":"The verification of Assumption (4.8) rests entirely on the two-level estimates (4.20), which are stated as assumptions in Lemma 4.2 and explicitly left unproven in Remark 3. Without (4.20), the stable decomposition (4.21), Assumption (4.8), and therefore the lower bound in Theorem 4.1 are not established. Consequently, the paper does not prove that the additive multigrid operator (4.6) satisfies the two-sided bound (3.24), and Corollary 3.1 cannot be invoked to deliver the advertised spectral equivalence for the negative-order preconditioner. The proposal in Remark 3 to derive (4.20) through the techniques of [9] is a strategy, not a proof; this is a load-bearing gap rather than a presentation issue.","section":"Section 4, Lemma 4.2 and Remark 3"},{"comment":"Even if (4.20) were available, the upper bound in (4.9) contains the explicit factor J, the number of grid levels. Since J grows as the mesh is refined, the asserted spectral equivalence constants are not independent of the discretization unless this factor is removed or otherwise shown to be harmless. The numerical experiments in Tables 1 and 3 keep J=4 in every run, so they do not exercise the level-dependence of the bound; reporting results with increasing J and a fixed coarsest mesh would be needed to support an h-uniform spectral equivalence claim.","section":"Theorem 4.1, Eq. (4.9)"}],"minor_comments":[{"comment":"The matrix entries (D_h)_{i,j} = -(φ_i^h, div ψ_i^h) appear to have an index typo; they should presumably involve ψ_j^h so that the matrix maps coefficient vectors on S_h to dual vectors on V_h.","section":"Remark 2, matrix realization"},{"comment":"There are several typographical errors, including 'charachteristic' in Section 4, 'adressed' in Section 4, 'Bogovski ˘ ı' in Section 3, and 'preconditioned conjugate method' in Section 5.2.","section":"Section 4 and Section 5"},{"comment":"The captions should state explicitly how the fixed number of levels J=4 is chosen relative to N; if the coarsest mesh changes with N, the experiments do not isolate the level-dependence of the method.","section":"Tables 1-3"},{"comment":"Reference [5] is cited as an arXiv preprint; the authors should update it to the published version if one exists, and should indicate more explicitly which parts of the present analysis rely on results from that preprint.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The author is unusually transparent about the missing two-level estimates, and the numerical experiments are consistent with the claims. Nevertheless, the central theoretical assertion of Section 4 is conditional, and the J-dependence in Theorem 4.1 needs attention. I recommend requiring either a proof of (4.20) or a revised statement that clearly labels the Section 4 construction as conditional, and additional experiments that vary J with a fixed coarsest mesh. If the missing proof can be supplied, this would be a solid contribution to solver theory for fractional-order interface problems."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does two things. First, it proves that B_s^h = ∇_h^* B_{div,h}^{1-s} ∇_h is spectrally equivalent to A_h^{-s} when B_{div,h}^{1-s} satisfies a two-sided bound relative to Λ_h^{-(1-s)}. That part is solid. The proof in Section 3 uses Jensen's operator inequality and the discrete Helmholtz decomposition, with explicit constants and a clean interpolation argument. This is the real contribution, and it is done well.\n\nSecond, it tries to supply B_{div,h}^{1-s} via an additive multigrid operator for fractional H(div). Here the analysis has a load-bearing gap. The lower bound in Theorem 4.1 requires Assumption 4.8, which is only established through Lemma 4.2. Lemma 4.2 depends on the two-level error estimates in (4.20)—bounds on the gradient and curl parts of (I−P^s_{k,k−1})τ. Those estimates are not proven. Remark 3 admits this explicitly and defers them to future work. So the abstract's claim that such operators are 'designed' goes beyond what is actually established.\n\nThe numerical experiments are consistent with the theory, but they fix J=4 and only increase N, so they cannot test the level-independence that the missing two-level estimate is supposed to provide. This is an omitted proof, not an internal contradiction. The paper is honest about the limitation, and Remark 3 suggests a plausible route through Bonito–Pasciak-style integral representations, but the gap is real.\n\nWhat follows is a conditional contribution. For someone working on trace-constrained multiphysics problems, Section 3 gives a useful framework that reduces negative-order fractional Laplacian preconditioning to positive-order fractional H(div) preconditioning. Section 4 provides a candidate operator with numerical evidence. If the two-level estimates are later proven, the paper becomes complete. Citation practice looks fine; the author builds on prior multigrid work for s≥0 and on standard H(div) multigrid, with no red flags.\n\nI would send this to a serious referee. The Section 3 result alone is publishable, and the Section 4 gap is explicitly flagged rather than hidden. A referee could reasonably ask for the gap to be filled or for the abstract to be softened, but the paper deserves the referee's time.","headline":"A clean, honest transfer argument for negative-order fractional Laplacian preconditioning, with the additive multigrid leg left conditional on unproven two-level estimates—worth refereeing despite the gap.","tokens_in":21996,"tokens_out":1882,"would_cite":false,"duration_ms":20357,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65F08","65N55","65N30","35R11"],"pacs":[],"model":"deepseek-v4-flash","headline":"Negative fractional Laplacians can be preconditioned through positive fractional H(div) operators via the discrete gradient.","keywords":["fractional Laplacian","negative fractional Sobolev spaces","auxiliary space preconditioner","additive multigrid","fractional H(div) operators","Raviart-Thomas elements","spectral equivalence","operator interpolation"],"falsifier":"Compute the ratios $\\|\\nabla_k u\\|^2/(h_{k-1}^{2s}(\\Lambda_k^s\\tau,\\tau))$ and $\\|q\\|/(h_{k-1}\\|\\tau\\|)$ over a sequence of uniformly refined meshes for $\\tau\\in(I-P^s_{k,k-1})V_k$ with the Helmholtz decomposition (4.19); if either ratio grows without bound as $h_k\\to0$, the stable decomposition in Lemma 4.2 fails and the additive multigrid preconditioner cannot be uniformly spectrally equivalent to $\\Lambda_h^{-s}$.","tokens_in":20935,"feed_emoji":"🧮","tokens_out":15655,"duration_ms":124019,"temperature":0.7,"pith_summary":"This paper attacks a problem that arises when coupled physics problems are solved with interface conditions posed in fractional Sobolev spaces: solving equations that involve the fractional Laplacian with a negative exponent, $(-\\Delta)^{-s}$ for $s\\in[0,1]$. Because the offending eigenvectors are the smooth ones, ordinary multigrid smoothing cannot handle the operator by itself; the author instead transfers the problem to a positive-order operator on the divergence side. The central result is that $B_s^h=\\nabla_h^* B_{\\mathrm{div},h}^{1-s}\\nabla_h$ is uniformly spectrally equivalent to $A_h^{-s}$ whenever $B_{\\mathrm{div},h}^{1-s}$ is a uniformly spectrally equivalent preconditioner for the fractional $H(\\operatorname{div})$ operator $\\Lambda_h^{-(1-s)}$, with constants $C_1\\beta^{2(1-s)}$ and $C_2$. The paper then constructs $B_{\\mathrm{div},h}^{s}$ as an additive multigrid operator built from vertex-patch smoothers, proves the upper bound and a conditional lower bound, and shows numerically that condition numbers stay flat. The stated caveat is that the lower-bound proof rests on two-level error estimates that the author leaves as future work.","feed_headline":"Negative fractional Laplacian gets a fast two-stage preconditioner","feed_subtitle":"It turns the negative exponent into a positive fractional H(div) problem, preconditioned by additive multigrid.","key_machinery":"The load-bearing object is the discrete Helmholtz decomposition $V_h=\\operatorname{curl}C_h\\oplus\\nabla_hS_h$ together with the operator $\\Lambda_h=I+\\nabla_h\\nabla_h^*$ on $V_h$. Since $\\Lambda_h$ commutes with both subspaces, its fractional power acts as the identity on $\\operatorname{curl}C_h$ and like $(I+A_h)^s$ on $\\nabla_hS_h$, so preconditioning $\\Lambda_h^s$ reduces to a scalar elliptic problem plus a trivial curl part. The auxiliary-space identity $B_s^h=\\nabla_h^*B_{\\mathrm{div},h}^{1-s}\\nabla_h$ transfers the negative fractional Laplacian into that positive fractional $H(\\operatorname{div})$ problem. The additive multigrid operator $\\sum_k R_k^sQ_k$ with vertex-patch smoothers supplies the $H(\\operatorname{div})$ factor, and Lemma 2.1, a Jensen operator inequality for contractions between different Hilbert spaces, carries the bounds from the endpoint cases to all $s\\in(0,1)$.","core_discovery":"On a shape-regular triangulation, $S_h$ is the space of piecewise polynomials, $V_h=\\mathrm{RT}_r(\\mathcal{T}_h)$ is the Raviart-Thomas space, and the discrete gradient $\\nabla_h:S_h\\to V_h$ is defined by $(\\nabla_h u,\\tau)=-(u,\\operatorname{div}\\tau)$. Theorem 3.2 proves the two-sided inequality $\\beta^{2(1-s)}A_h^s\\le\\nabla_h^*\\Lambda_h^{-(1-s)}\\nabla_h\\le A_h^s$, and Corollary 3.1 converts any $B_{\\mathrm{div},h}^{1-s}$ satisfying (3.24) into $C_1\\beta^{2(1-s)}A_h^s\\le B_s^h\\le C_2A_h^s$. The proof interpolates between the endpoint cases $s=0$ and $s=1$ through Jensen's operator inequality. For the $H(\\operatorname{div})$ factor, the additive multigrid operator $B_{\\mathrm{div},h}^s=\\sum_{k=1}^J R_k^s Q_k$ uses vertex-patch additive Schwarz smoothers $R_k^s=\\sum_\\nu\\Lambda_{k,\\nu}^{-s}Q_{k,\\nu}$, and Theorem 4.1 gives spectral equivalence to $\\Lambda_h^{-s}$ under the smoother assumptions (4.7) and (4.8). The lower-bound half is conditional: Lemma 4.2 reduces (4.8) to the two-level estimates (4.20), and Remark 3 explicitly states that proving them is left as future work.","pith_inferences":["A sharper stress test would fix $J$ and push refinements deeper; the theorem's upper bound scales linearly in $J$, so condition numbers should grow at most linearly in $J$, and the flat table with $J=4$ cannot distinguish the proven upper bound from a failing lower bound.","The integral-representation route sketched in Remark 3, if completed, would tie the fractional H(div) two-level estimates to standard estimates for $(I-t^2\\Delta)^{-1}$ and would identify the exact domain-regularity condition needed, likely making the proof transfer to nonuniform meshes.","Because the argument only uses the inf-sup condition (3.15), a commuting Helmholtz decomposition, and stable vertex-patch decompositions, the same two-stage construction should carry over to higher-order Raviart-Thomas spaces and to three dimensions.","For the motivating trace problem, the construction would replace a fractional operator on the interface by standard H(div) multigrid on the interface mesh; the remaining difficulty would then be geometric approximation of the trace space by coarse interface meshes."],"forward_implications":["For any $s\\in[0,1]$, a preconditioner $B_{\\mathrm{div},h}^{1-s}$ satisfying (3.24) lifts to a preconditioner for $A_h^{-s}$ with h-independent constants $C_1\\beta^{2(1-s)}$ and $C_2$, so the negative-order problem inherits the quality of the positive-order H(div) preconditioner.","If the two-level estimates (4.20) hold, the additive multigrid operator $B_{\\mathrm{div},h}^s$ is uniformly spectrally equivalent to $\\Lambda_h^{-s}$ for every $s\\in[0,1]$ by interpolation between the cases $s=0$ and $s=1$.","The numerical experiments keep conjugate-gradient iteration counts and estimated condition numbers flat across four levels of refinement for both $\\Lambda_h^s$ and $A_h^{-s}$, consistent with the proven upper bound and the conditional lower bound.","Applied to the trace-constrained system (1.2), the new operator supplies the fractional block of an ideal block-diagonal preconditioner without requiring a global fractional power of the Laplacian."],"supporting_citations":[{"why":"supplies the additive multigrid construction for fractional Sobolev spaces that this paper extends from nonnegative to negative exponents","marker":"[5]"},{"why":"introduces the auxiliary space preconditioning framework that motivates the transfer operator $\\nabla_h$","marker":"[32]"},{"why":"gives the operator preconditioning criterion that fixes which norm the spectral equivalence must target","marker":"[28]"},{"why":"provides the left-inverse $T$ of the gradient with the mapping properties used in the lower bound of Theorem 3.1","marker":"[15]"},{"why":"supplies the stable vertex-patch decompositions and the H(div) multigrid analysis behind the smoother","marker":"[2]"},{"why":"supplies the H(div) and H(curl) multigrid theory and the discrete Helmholtz decomposition facts used in the two-level argument","marker":"[3]"},{"why":"gives the fractional-power error analysis proposed as the route to prove the unproven two-level estimates (4.20)","marker":"[9]"}],"fun_headline_variants":["Fast two-stage preconditioner for negative fractional Laplacian","Auxiliary space trick turns negative Laplacian into positive H(div)","Negative fractional Laplacian preconditioned via positive H(div)","Two-stage preconditioner trick for negative fractional Laplacian"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the pair of two-level error estimates (4.20): for an error $\\tau\\in(I-P^s_{k,k-1})V_k$ with discrete Helmholtz decomposition $\\tau=\\nabla_k u+\\operatorname{curl}q$, one must have $\\|\\nabla_k u\\|^2\\le c h_{k-1}^{2s}(\\Lambda_k^s\\tau,\\tau)$ and $\\|q\\|\\le ch_{k-1}\\|\\tau\\|$. The paper uses these estimates in Lemma 4.2 to verify the stable decomposition that powers the lower bound, and Remark 3 leaves their proof to future work, so the spectral equivalence of the multigrid preconditioner is conditional on them.","fun_headline_variants_meta":{"raw":{"variants":["Fast two-stage preconditioner for negative fractional Laplacian","Auxiliary space trick turns negative Laplacian into positive H(div)","Negative fractional Laplacian preconditioned via positive H(div)","Two-stage preconditioner trick for negative fractional Laplacian"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.002431,"raw_usage":{"total_tokens":9398,"prompt_tokens":1052,"completion_tokens":8346,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":668,"completion_tokens_details":{"reasoning_tokens":8283}},"tokens_in":668,"tokens_out":8346,"duration_ms":66628,"temperature":1.0,"reasoning_tokens":8283,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:40:25.691102+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the ratios $\\|\\nabla_k u\\|^2/(h_{k-1}^{2s}(\\Lambda_k^s\\tau,\\tau))$ and $\\|q\\|/(h_{k-1}\\|\\tau\\|)$ over a sequence of uniformly refined meshes for $\\tau\\in(I-P^s_{k,k-1})V_k$ with the Helmholtz decomposition (4.19); if either ratio grows without bound as $h_k\\to0$, the stable decomposition in Lemma 4.2 fails and the additive multigrid preconditioner cannot be uniformly spectrally equivalent to $\\Lambda_h^{-s}$.","supporting_citations":[{"cited_title":"Multigrid Methods for Discrete Fractional Sobolev Spaces","cited_arxiv_id":"1806.00222","evidence_quote":"supplies the additive multigrid construction for fractional Sobolev spaces that this paper extends from nonnegative to negative exponents"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"introduces the auxiliary space preconditioning framework that motivates the transfer operator $\\nabla_h$"},{"cited_title":"Mardal and R","cited_arxiv_id":null,"evidence_quote":"gives the operator preconditioning criterion that fixes which norm the spectral equivalence must target"},{"cited_title":"Costabel and A","cited_arxiv_id":null,"evidence_quote":"provides the left-inverse $T$ of the gradient with the mapping properties used in the lower bound of Theorem 3.1"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the stable vertex-patch decompositions and the H(div) multigrid analysis behind the smoother"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the H(div) and H(curl) multigrid theory and the discrete Helmholtz decomposition facts used in the two-level argument"},{"cited_title":"Bonito and J","cited_arxiv_id":null,"evidence_quote":"gives the fractional-power error analysis proposed as the route to prove the unproven two-level estimates (4.20)"}],"review_version":1}