{"id":"16e495d8-eebb-40d1-af96-3905ce736e00","arxiv_id":"2505.17191","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An adaptive-rank, greedy-sampling semi-Lagrangian solver with a macroscopic conservation correction is developed for the BGK equation and shown to be accurate, conservative, and conditionally asymptotic-preserving.","lead":"This paper presents a new low-rank numerical method for the multi-scale BGK equation of gas dynamics, using greedy sampling to update only key parts of the solution and a correction step to preserve mass, momentum, and energy. It matters because it offers a way to simulate rarefied gas flows across fluid and kinetic regimes with large time steps and reduced storage.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.6's proof is circular: after LoMaC convergence, the quantity being bounded equals the provisional low-rank deviation, and the E-terms in (3.29) telescope to the LoMaC adjustment, not the accumulated truncation error relative to the full-grid solution.","rationale":"The paper's strongest theoretical contribution is the conditional asymptotic-preserving property, and that property rests on Theorem 3.6 and Corollary 3.7. The reader correctly identified the conditionality of the result: the truncation error must be O(ε), which may fail because local Maxwellians are not generally low-rank. My concern is more specific and more damaging to the proof: the algebra in Theorem 3.6 does not separate the low-rank truncation error from the quantity being bounded. After LoMaC convergence, the deviation F − M_U(F) is identically equal to the provisional low-rank deviation F* − M_U(F*); therefore the proof's first term cannot be declared O(ε) by appealing to the full-grid AA property. The E-terms in (3.29) telescope to the LoMaC adjustment M_U(F*) − M_U, not to the CUR/SVD errors introduced by the S-stage and L-stage compressions. In particular, the S-stage truncation error discussed in (3.31) is not represented in the theorem's bound, and the authors themselves state that its accumulation is difficult to analyze after SVD truncation destroys the CUR interpolation property. This leaves the core asymptotic claim weaker than the paper states: Corollary 3.7's condition is imposed on the wrong terms, and the proof needs either a new decomposition that tracks the actual truncation error through both stages or an explicit additional condition on that error. I want to credit the parts that do hold independently: the LoMaC correction format enforces conservation if the macroscopic system is solved accurately, the numerical experiments are extensive and internally consistent, and the convergence tables and rank data support the practical claims. The issue here is confined to the theoretical AP analysis, but it is load-bearing because the abstract and introduction advertise the conditional AP property as a central contribution. A conditional-accept verdict remains appropriate provided the authors correct Theorem 3.6, revise the condition in Corollary 3.7 to include the missing truncation terms, and ideally add a numerical check of the modified condition. The verdict is therefore CONDITIONAL rather than ACCEPT or REJECT: the method may well be sound, but the published proof of its central asymptotic property is not yet valid as written.","tokens_in":21315,"tokens_out":14549,"duration_ms":113956,"concrete_test":"Re-derive (3.27)-(3.29) under the exact converged condition ∫ F^{(k)} Φ dv = U^{(k)}. Show that F^{(k)} − M_U(F^{(k)}) = F^{(k),*}_{...,r2} − M_U(F^{(k),*}_{...,r2}) and that the second and third terms in (3.27) cancel identically; if so, the stated bound is not a consequence of the decomposition. Then run the smooth test of Example 4.1 with ϵ=10^-6 and deliberately loose truncation tolerances (e.g., εC=εS=10^-3), and compare ||F − M_U(F)|| against ε + ||E^{εs}|| + ||E^{εc}||. If the left-hand side is not bounded by the right-hand side, Theorem 3.6 as stated fails in a numerically accessible regime, and the missing term is the S-stage truncation error.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Under the converged LoMaC correction, the moments of the corrected solution satisfy ∫ F^{(k)} Φ dv = U^{(k)}, so M_U(F^{(k)}) = M_{U^{(k)}} (with the moments of F* understood in the Maxwellian notation). Equation (3.23) then gives exactly F^{(k)} − M_U(F^{(k)}) = F^{(k),*}_{...,r2} − M_U(F^{(k),*}_{...,r2}). Thus the first term in the decomposition (3.27) is the very quantity Theorem 3.6 is meant to bound. The sentence immediately before (3.28), \"Since the initial data is well-prepared and the local solver is AA, we immediately have ... ≲ ε,\" therefore assumes the conclusion for the low-rank solution rather than proving it; the full-grid AA property of [20] applies to the untruncated scheme, not to the truncated F*. Moreover, the two E-terms in (3.29) telescope to F^{(k),*}_{...,r2} − F^{(k)} = M_U(F*) − M_U, i.e., the LoMaC adjustment only. They are not the accumulated CUR/SVD error of the S-stage and L-stage relative to the full-grid solution: the S-stage truncation error is absent from the stated bound, and the authors concede after (3.31) that it is difficult to control. Consequently, Corollary 3.7 imposes its O(ε) condition on the wrong quantities; the RHS of Theorem 3.6 can be small while F^{(k)} − M_U(F^{(k)}) is large whenever the provisional low-rank solution is far from its own Maxwellian. This is an internal gap in the asymptotic analysis, not merely an unfavorable choice of tolerances.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an adaptive-rank numerical method for the multi-scale BGK equation. The method combines a greedy adaptive cross approximation (ACA) with SVD truncation, a local semi-Lagrangian finite-difference solver, and a locally macroscopic conservative (LoMaC) correction that enforces mass, momentum, and energy conservation by implicitly coupling the kinetic solution to a macroscopic moment system. High-order temporal accuracy is obtained with stiffly accurate DIRK methods, and the resulting nonlinear systems are solved with a Jacobian-free Newton-Krylov method. The main theoretical claims are that the LoMaC correction preserves the asymptotic-preserving (AP) and asymptotically accurate (AA) properties of the full-grid scheme, and that the low-rank scheme is conditionally AP provided certain truncation errors are O(epsilon) (Theorem 3.6 and Corollary 3.7). Numerical experiments report third-order temporal and spatial convergence for a range of Knudsen numbers, machine-precision conservation, and robust behavior on a Riemann problem and a mixed-regime problem with spatially varying Knudsen number.","tokens_in":21673,"tokens_out":4542,"duration_ms":34597,"significance":"If the theoretical claims are sound, the paper makes a useful contribution: it extends the semi-Lagrangian adaptive-rank framework to collisional kinetic equations without forming explicit low-rank decompositions of the local Maxwellian, and it demonstrates strict conservation of all three collision invariants. The numerical evidence is strong: the convergence tables are clean, conservation is attained to machine precision across all tested regimes, and the mixed-regime experiments show stable behavior over several orders of magnitude of the Knudsen number. The authors are also transparent about limitations, including the conditional nature of the AP result and the difficulty of controlling accumulated truncation error. However, the central asymptotic analysis, specifically the proof of Theorem 3.6, contains a load-bearing gap that affects Corollary 3.7 and therefore the paper's main theoretical claim. The numerical experiments are convincing as algorithmic demonstrations, but they do not currently close that gap.","major_comments":[{"comment":"The proof of Theorem 3.6 is circular and does not establish the stated bound. After the LoMaC correction (3.23), the moments of F^(k) are enforced to equal U^(k), so M_U(F^(k)) = M_{U^(k)} and (3.23) gives F^(k) - M_U(F^(k)) = F^(k),*_{#,k2->r2} - M_U(F^(k),*_{#,k2->r2}). The first term in the decomposition (3.27) is therefore exactly the quantity the theorem is supposed to bound. The sentence preceding (3.28), which asserts that 'the initial data is well-prepared and the local solver is AA' implies the bound, applies the full-grid AA result of [20] to the truncated provisional solution, for which that result is not proven. In addition, the two E-terms in (3.29) sum to F^(k),*_{#,k2->r2} - F^(k), which by (3.23) equals the LoMaC adjustment M_U(F^(k),*_{#,k2->r2}) - M_{U^(k)}; they do not include the S-stage truncation error, and the discussion after (3.31) concedes that this accumulation is difficult to control. Consequently, Corollary 3.7 imposes its O(epsilon) condition on quantities that can be small while F^(k) - M_U(F^(k)) is large, so the conditional AP/AA property is not established by the present argument.","section":"§3.5, Theorem 3.6"},{"comment":"Even if Theorem 3.6 were repaired, the conditional AP property as stated is substantially weaker than the abstract suggests. The condition ||E_eps_s(F^(k)_{#,k2})|| + ||E_eps_c(F^(k))|| lesssim epsilon forces the truncation tolerances to scale with the Knudsen number, and Remark 3.8 admits that this may require large ranks and undermine efficiency. The generic fluid limit has a local Maxwellian that is not low-rank in the x-v tensor format except under rank-degeneracy or special assumptions. The numerical tests in Section 4 use fixed tolerances (for example, epsilon_C = 1e-9 and epsilon_S = 1e-8 in Example 4.1, and much looser tolerances in Example 4.2) and do not demonstrate that the Corollary's condition is satisfied. The manuscript should clarify whether Corollary 3.7 is intended as a proof of a practical AP property or as a consistency condition that practical tolerance choices must satisfy.","section":"Corollary 3.7 and Remark 3.8"},{"comment":"The AP/AA analysis assumes well-prepared initial data (Definition 3.3) and exact spatial interpolation at the characteristic feet, as stated in Theorem 3.4. The numerical experiments use WENO spatial reconstruction and, in Example 4.2, discontinuous Riemann initial data, so the theory does not cover the regimes in which the method is actually tested. The authors are transparent about these restrictions, but the manuscript should state explicitly that the numerical evidence is heuristic support for the method rather than verification of the conditional AP theorem in the tested configurations.","section":"Theorem 3.4 and numerical tests"}],"minor_comments":[{"comment":"The expression 'eR\nR± ·dv' in the text after equation (3.14) appears to be a rendering artifact; it should read as an integral over the half-velocity domains, e.g., ∫_{R±} · dv.","section":"§3.3.1"},{"comment":"In the sentence 'Our results indicate that smaller values of epsilon generally lead to an SLightly higher number of iterations', 'SLightly' should be 'slightly'.","section":"Example 4.2"},{"comment":"Remark 3.9 refers to 'Theorem 3.7', but the conditional AA property is stated in Corollary 3.7; the cross-reference should be corrected.","section":"Remark 3.9"},{"comment":"The notation M_U(F^(k)) versus M_{U^(k)} is used without explicit definition of the latter; since the distinction between 'Maxwellian with moments of the provisional solution' and 'Maxwellian with the enforced macroscopic moments' is central to the LoMaC correction, a short notational clarification would improve readability.","section":"Notation in Section 3.3 and 3.5"}],"recommendation":"major_revision","confidential_remarks":"For the editor: The numerical work is strong and the algorithmic idea is interesting, but the asymptotic analysis needs substantive revision before the paper can be accepted. In particular, Theorem 3.6 has a genuine circularity, and Corollary 3.7's conditional statement is currently more an assumption than a proven property. I believe the authors can repair this, perhaps by proving a different bound for the deviation from the local Maxwellian or by reformulating the conditional AP claim, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Sarah,\n\nPunchline first: the method in this paper is a real step forward, and the numerics are convincing. The asymptotic analysis, however, has a circular proof in Theorem 3.6, and that needs to be fixed before the conditional AP claim is used as a selling point.\n\nWhat's genuinely new: the authors extend their SLAR framework to the BGK equation using greedy ACA plus SVD truncation, which sidesteps the need to form low-rank decompositions of the local Maxwellian. The LoMaC correction is a natural extension of their prior mass-conservation trick to all three conserved moments, and the implicit coupling to a macroscopic system via JFNK is well executed. The experiments are clean: third-order temporal and spatial convergence across a range of Knudsen numbers, conservation to machine precision, and a mixed-regime test with epsilon varying by six orders of magnitude. The authors also flag the rank-degeneracy caveat honestly in Remark 3.8, so they are not overselling the theory.\n\nThe soft spot is in Theorem 3.6. Under the converged LoMaC correction, the moments of F(k) match U(k), so the quantity ||F(k) - M_U(F(k))|| equals ||F(k),* - M_U(F(k),*)||. That is exactly the deviation of the provisional low-rank solution from its own Maxwellian. The proof's line 'Since the initial data is well-prepared and the local solver is AA' assumes this deviation is O(epsilon); it does not prove it for the truncated solution. The E-terms in (3.29) also telescope to the LoMaC adjustment, not to the accumulated CUR/SVD error relative to the full-grid solution. So Corollary 3.7 imposes its condition on the wrong quantities. This is a real gap, though it does not invalidate the numerical claims.\n\nMinor points: no code is provided, so independent reproducibility is limited, and the analysis assumes exact spatial interpolation and well-prepared data, which the authors acknowledge.\n\nWho this is for: numerical analysts working on low-rank kinetic solvers. The method itself is worth a serious referee; the theory section needs major revision. I'd recommend sending it out, with the requirement that the authors either patch the proof with a genuine error-accumulation argument or explicitly downgrade the AP statement to a formal heuristic.","headline":"Promising low-rank BGK solver with strong numerics; the conditional AP theorem's proof is circular and needs rework before the theoretical claims are cited.","tokens_in":22213,"tokens_out":3304,"would_cite":true,"duration_ms":22985,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65M06","65F30","35Q20","65L06"],"pacs":[],"model":"deepseek-v4-flash","headline":"An adaptive-rank BGK scheme conserves mass, momentum, and energy to machine precision while capturing fluid-kinetic regimes.","keywords":["BGK equation","adaptive-rank methods","greedy cross approximation","semi-Lagrangian methods","asymptotic-preserving schemes","macroscopic conservation","Knudsen number","Newton-Krylov"],"falsifier":"On a benchmark whose fluid limit has small numerical rank (so Corollary 3.7's condition is met), set the truncation tolerances to $\\epsilon_c=\\epsilon_s=\\epsilon$, take one time step at CFL 4, and measure $\\lVert F^{n+1}-M_U(F^{n+1})\\rVert_\\infty$ as $\\epsilon\\to 0$. If the deviation does not decay like $O(\\epsilon)$, the conditional asymptotic-preserving claim fails even where its assumptions hold.","tokens_in":21099,"feed_emoji":"⚛️","tokens_out":9325,"duration_ms":74138,"temperature":0.7,"pith_summary":"This paper sets out to show that the BGK relaxation model for rarefied gases can be solved by an adaptive-rank method that never builds a low-rank decomposition of the local Maxwellian. The scheme uses greedy adaptive cross approximation to select important rows and columns of the distribution-function matrix, updates only those entries with a local semi-Lagrangian solver, and reconstructs the rest of the matrix with a CUR decomposition followed by SVD truncation. A locally macroscopic conservative correction then couples the low-rank kinetic solution to an implicit macroscopic moment system, so mass, momentum, and energy are enforced to machine precision. The result is a multi-scale solver that is third-order accurate, captures shocks and mixed fluid-kinetic regimes, and remains asymptotic-preserving in the fluid limit provided the truncation error is $O(\\epsilon)$.","feed_headline":"Adaptive-rank BGK solver conserves moments to machine precision","feed_subtitle":"Greedy sampling plus a coupled macroscopic correction handles shocks across six orders of Knudsen number.","key_machinery":"The load-bearing object is the ACA+SVD compression routine, in which adaptive cross approximation builds a CUR decomposition $A\\approx A(:,J)A(I,J)^{-1}A(I,:)$ by greedily choosing pivots from the residual, followed by an SVD truncation that removes redundant modes. Applied once to the semi-Lagrangian characteristic update and once to the BGK relaxation update, the routine updates only the selected rows and columns, so the local Maxwellian is evaluated pointwise at sample locations rather than decomposed. The second carrying mechanism is the locally macroscopic conservative (LoMaC) correction, an implicit coupling of the low-rank kinetic solution to a macroscopic moment system whose flux is closed dynamically by the kinetic solution; the resulting nonlinear system is solved by a Jacobian-free Newton-Krylov method with the provisional solution as initial guess.","core_discovery":"The central claim is that greedy sampling plus SVD truncation can carry the full multi-scale behavior of the BGK equation without explicit low-rank decompositions of its nonlinear collision term. After a provisional nonconservative low-rank update $F^{n+1,\\star}$, the LoMaC correction imposes conservation through $F^{n+1}=F^{n+1,\\star}-M_{U(F^{n+1,\\star})}+M_{U^{n+1}}$, where $U^{n+1}$ solves an implicit, stiffly accurate DIRK discretization of the macroscopic flux system with a kinetic-flux-splitting space discretization. The asymptotic analysis proves that the full-grid scheme's asymptotic-preserving and asymptotically accurate behavior survives the correction (Theorem 3.5), and that the low-rank scheme is conditionally asymptotically accurate when the combined CUR and SVD truncation errors are $O(\\epsilon)$ (Corollary 3.7). Numerical studies show third-order temporal and spatial convergence for Knudsen numbers from $10^{-6}$ to $10^{-2}$, correct shock and rarefaction structures, and stable behavior on a mixed-regime problem with space-varying Knudsen number.","pith_inferences":["If the distribution is written multiplicatively as $f=M_{U(f)}g$, the limiting fluid solution has rank one, which would likely turn the conditional asymptotic-preserving guarantee into an unconditional one; the authors explicitly leave this variant for future work.","The pointwise-evaluation trick for the local Maxwellian may transfer to other stiff kinetic operators, such as Fokker-Planck or full Boltzmann collisions, wherever the nonlinear term can be evaluated locally without low-rank assembly.","A scaling study of the observed rank versus $\\epsilon$ at fixed error tolerance would show where the method's efficiency advantage starts to degrade, since a generic local Maxwellian is not low rank.","The moment correction behaves like a dynamic closure; it would be worth testing whether it also damps spurious oscillations of the low-rank solution in strongly non-equilibrium regions."],"forward_implications":["Mass, momentum, and energy are conserved to machine precision despite low-rank compression.","Large CFL numbers are feasible: CFL 4 is used throughout the numerical tests, with small Newton and Krylov iteration counts and no preconditioning at moderate CFL.","The scheme captures shocks, contact discontinuities, and rarefactions across Knudsen numbers from $10^{-6}$ to $10^{-2}$.","For mixed regimes in which the Knudsen number varies over six orders of magnitude, the solution rank stays modest, using roughly 15–23% of full-grid storage.","With truncation tolerances chosen $O(\\epsilon)$, the low-rank scheme inherits the asymptotic-preserving property of the underlying full-grid semi-Lagrangian method."],"supporting_citations":[{"why":"Supplies the semi-Lagrangian adaptive-rank framework and the ACA+SVD compression core that this paper extends to the BGK equation.","marker":"[1]"},{"why":"Provides the fifth-order semi-Lagrangian finite-difference solver with WENO reconstruction used as the local kinetic update.","marker":"[19]"},{"why":"Establishes the asymptotic-preserving and asymptotically accurate properties of the full-grid semi-Lagrangian DIRK discretization that the analysis builds on.","marker":"[20]"},{"why":"Introduces the locally macroscopic conservation correction for low-rank kinetic solvers that is generalized here to momentum and energy.","marker":"[13]"},{"why":"Supplies the adaptive cross approximation algorithm and its recursive rank-one updates used for the greedy sampling.","marker":"[15]"},{"why":"Supplies the kinetic flux vector splitting used to build upwind numerical fluxes for the macroscopic system.","marker":"[27]"},{"why":"Supplies the Jacobian-free Newton-Krylov methodology used to solve the implicit macroscopic system without forming the Jacobian.","marker":"[30]"},{"why":"Supplies the mixed-regime benchmark with a spatially varying Knudsen number that tests the method across six orders of magnitude.","marker":"[35]"}],"fun_headline_variants":["Greedy low-rank BGK solver preserves moments across scales","Adaptive-rank BGK method captures shocks with exact conservation","Low-rank BGK via greedy sampling handles all Knudsen regimes","CUR-based BGK integrator conserves moments without low-rank splits","Adaptive rank BGK: shocks and moments correct from 1e-6 to 1e-2"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the combined CUR and SVD truncation error is $O(\\epsilon)$ as the Knudsen number goes to zero, which is not automatic because a generic local Maxwellian does not have low rank.","fun_headline_variants_meta":{"raw":{"variants":["Greedy low-rank BGK solver preserves moments across scales","Adaptive-rank BGK method captures shocks with exact conservation","Low-rank BGK via greedy sampling handles all Knudsen regimes","CUR-based BGK integrator conserves moments without low-rank splits","Adaptive rank BGK: shocks and moments correct from 1e-6 to 1e-2"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000179,"raw_usage":{"total_tokens":1368,"prompt_tokens":1083,"completion_tokens":285,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":699,"completion_tokens_details":{"reasoning_tokens":201}},"tokens_in":699,"tokens_out":285,"duration_ms":2514,"temperature":1.0,"reasoning_tokens":201,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:50:27.874581+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a benchmark whose fluid limit has small numerical rank (so Corollary 3.7's condition is met), set the truncation tolerances to $\\epsilon_c=\\epsilon_s=\\epsilon$, take one time step at CFL 4, and measure $\\lVert F^{n+1}-M_U(F^{n+1})\\rVert_\\infty$ as $\\epsilon\\to 0$. If the deviation does not decay like $O(\\epsilon)$, the conditional asymptotic-preserving claim fails even where its assumptions hold.","supporting_citations":[{"cited_title":"Zheng, D","cited_arxiv_id":null,"evidence_quote":"Supplies the semi-Lagrangian adaptive-rank framework and the ACA+SVD compression core that this paper extends to the BGK equation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the fifth-order semi-Lagrangian finite-difference solver with WENO reconstruction used as the local kinetic update."},{"cited_title":"Ding, J.-M","cited_arxiv_id":null,"evidence_quote":"Establishes the asymptotic-preserving and asymptotically accurate properties of the full-grid semi-Lagrangian DIRK discretization that the analysis builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the locally macroscopic conservation correction for low-rank kinetic solvers that is generalized here to momentum and energy."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the kinetic flux vector splitting used to build upwind numerical fluxes for the macroscopic system."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Jacobian-free Newton-Krylov methodology used to solve the implicit macroscopic system without forming the Jacobian."},{"cited_title":"Xiong, J","cited_arxiv_id":null,"evidence_quote":"Supplies the mixed-regime benchmark with a spatially varying Knudsen number that tests the method across six orders of magnitude."}],"review_version":1}