{"id":"f675d9ce-f035-4066-bfb0-68ed0bee3ab4","arxiv_id":"1908.03688","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A Lagrangian-frame DMD that includes characteristic-line positions as Koopman observables yields accurate reduced-order models for shock-free advection-dominated PDEs.","lead":"This paper proposes a Lagrangian version of dynamic mode decomposition (DMD) that uses particle paths and solution values together as observables. It shows, on 1D examples, that this recovers accurate reduced-order models for advection-dominated flows where standard DMD and POD fail.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lagrangian DMD's central claim rests on characteristic trajectories that all numerical tests obtain from a Lagrangian HFM; the Eulerian-only-data route is never demonstrated for DMD.","rationale":"The reader's verdict is CONDITIONAL and identifies the same general gap: accurate characteristic trajectories are assumed available, and the Eulerian-only route is untested. I agree that this is the main weakness. My partial disagreement is on one point of framing: Section 3.1 does not actually describe an Eulerian-only route for Lagrangian DMD; it describes a complication for Lagrangian POD and points to [30, Sec. 3.3] for an optimal basis. No corresponding data-recovery procedure for the DMD observables is specified, so the gap is slightly larger than the reader's wording suggests. I do not see an internal inconsistency in the shock-free numerical demonstrations, and the paper candidly limits its conclusions in Section 5. The numerical tests are consistent with the claim that, when exact Lagrangian trajectories are provided, Lagrangian DMD extrapolates well on these shock-free problems. The unresolved question is whether those trajectories can be obtained in the Eulerian-data setting that motivates DMD. That question is testable directly, and the test proposed above would settle it. Since the reader already conditioned acceptance on related concerns, my stress-test does not change the verdict.","tokens_in":14043,"tokens_out":4399,"duration_ms":53313,"concrete_test":"Run the Section 4.2 advection-diffusion test using only Eulerian HFM snapshots u^n from (2.3), and estimate characteristic trajectories x^n by particle tracking: integrate dX/dt = f(u(X,t)) with interpolation of the Eulerian snapshots (or by the Lagrangian-basis construction referenced in Section 3.1 from [30, Sec. 3.3]). Then apply Algorithm 3.1 to y^n = [x^n; u^n] and compare the extrapolated solution for t > 0.25 with the reference and with the reported Lagrangian-HFM-based results. If the Eulerian-only Lagrangian DMD error is comparable (within one order of magnitude) to the published error, the central claim extends to the standard data regime; if not, the method is restricted to cases where a Lagrangian HFM is already available.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that Algorithm 3.1, with observables y^n = [x^n; u^n] from (3.11), builds a working ROM for advection-dominated problems. For that claim to hold in the usual DMD setting, the characteristic positions x^n must be recoverable from the available training data. The paper's numerical tests, however, effectively supply x^n from a Lagrangian high-fidelity solve: Section 4 states that reference solutions are computed with the Eulerian scheme (2.3), but the Lagrangian trajectories used as observables are not themselves obtained from those Eulerian snapshots by any demonstrated procedure. Section 3.1 mentions that, for Eulerian-only data, an optimal Lagrangian basis can be constructed following [30, Sec. 3.3], but that remark is about Lagrangian POD and no analogous construction or test is given for Lagrangian DMD. Thus, for an end user who has only Eulerian snapshots (the standard DMD data format), the paper does not show that the required x^n can be produced accurately. If the reconstructed trajectories are inaccurate, errors propagate directly into the DMD fit and the extrapolating predictions. The shock-free restriction is explicitly acknowledged in Section 5, so it is not a hidden flaw; the unvalidated dependence on Lagrangian-solver-generated trajectories is the load-bearing soft spot.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a Lagrangian reformulation of dynamic mode decomposition for reduced-order modeling of advection-dominated PDEs. It first demonstrates with a linear advection example that standard Eulerian POD and DMD generate global SVD bases that cannot represent a translating pulse, leading to inaccurate extrapolation. It then introduces Algorithm 3.1, which applies the standard DMD procedure to the augmented observable y^n = [x^n; u^n], where x^n denotes characteristic-line positions, and reconstructs u from the predicted observable. Numerical experiments on linear advection, linear advection-diffusion, inviscid and viscous Burgers equations, and a level-set formulation are presented; all reported tests are shock-free. In these tests, Lagrangian DMD and Lagrangian POD capture the solution beyond the training window t > 0.25, with DMD being cheaper because it is iteration-free. An error bound from a previous paper by the same authors is quoted and plotted.","tokens_in":14303,"tokens_out":10552,"duration_ms":105952,"significance":"Should the method hold, it offers a simple physics-aware extension of DMD that is equation-free and iteration-free and that directly addresses the translation failure of Eulerian SVD methods. The choice of [x; u] as observables is physically motivated and is not tuned to the test outputs. The numerical comparisons against conventional Eulerian DMD and POD are clean and convincing for the shock-free cases. The main caveat is that the method's success depends on the availability of accurate characteristic trajectories; this dependence is acknowledged only indirectly and is not validated in the standard Eulerian-snapshot data regime. Given the explicit shock-free limitation and the unvalidated data pipeline, the contribution is promising but not yet established at the claimed scope.","major_comments":[{"comment":"The numerical validation supplies the characteristic positions x^n from a separate Lagrangian high-fidelity solve, as the listing of 'Lagrangian HFM computational time' in Table 1 indicates, rather than deriving them from the Eulerian reference snapshots produced by (2.3). For an end user with Eulerian-only data, the route described in Section 3.1 is stated only for Lagrangian POD (following [30, Sec. 3.3]) and is neither formulated nor tested for Lagrangian DMD. Because Algorithm 3.1 requires x^n as an observable, this is a load-bearing gap: if characteristic trajectories cannot be recovered accurately from the available data, the method's advantage disappears. The authors should either demonstrate trajectory recovery from Eulerian snapshots and use it in the numerical tests, or explicitly restrict the paper's claim to settings in which Lagrangian trajectory data are available.","section":"Section 4 / Table 1"},{"comment":"The error bound used in every numerical example is quoted from [20], but epsilon_m is not defined in this manuscript and the theorem is stated without proof or a precise pointer to the statement in [20]. Since the figures plot this bound as an estimate of the observable error, a reader cannot verify whether the plotted curve is actually the right-hand side of (4.2). The authors should either include the definition of epsilon_m and the main steps of the proof, or remove the bound from the figures and state the error analysis as a reference.","section":"Theorem 4.1"}],"minor_comments":[{"comment":"The matrix X' should be Y2; as written, the low-rank approximation K-tilde is not defined in terms of the observable data matrices introduced in Step 0.","section":"Algorithm 3.1, Step 2"},{"comment":"The prediction formulas use Lambda^{n+1} with b = Phi^{-1} y^1 (or u^0), but the standard DMD convention with snapshots indexed from 1 gives y^n = Phi Lambda^{n-1} b. Please clarify the indexing so the extrapolation formula is unambiguous.","section":"Eqs. (2.17) and (3.9)"},{"comment":"The caption says 'linear advection equation,' but Section 4.2 concerns the linear advection-diffusion equation; please correct.","section":"Figure 7 caption"},{"comment":"The conclusion explicitly states that all tests are shock-free, but the abstract and introduction do not carry this qualification; readers may overinterpret the scope. Please state the shock-free limitation prominently in the abstract.","section":"Abstract and Introduction"},{"comment":"The phrases 'optimal Lagrangian basis' and 'optimal choice of observable functions' are not backed by an optimality criterion; 'well-motivated' or 'physics-informed' would be more accurate.","section":"Sections 3.1 and 3.2"},{"comment":"The level-set DMD experiment would benefit from a precise statement of the observable matrix, the number of snapshots, the SVD rank, and the reconstruction procedure to be reproducible.","section":"Appendix A"},{"comment":"There are several typos, e.g., 'explaination' and 'physic-aware'; a careful proofreading pass is needed.","section":"Section 1"}],"recommendation":"major_revision","confidential_remarks":"The error bound in Theorem 4.1 comes from a closely related preprint by the same authors; please ensure that the present paper does not present this as a new contribution and that the definition in [20] is accessible to readers. The absence of an Eulerian-only-data validation is the main risk to the central claim. A comparison with other DMD variants that address advection, such as time-delay or extended DMD, would help calibrate the contribution, though the paper's baseline comparison with conventional Eulerian DMD/POD is a reasonable starting point."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The right way to read this paper is as a useful incremental extension of Lagrangian POD to DMD, not as a complete solution to advection-dominated ROMs. The core demonstration—that using [x;u] as observables fixes the translational failure of SVD-based methods—works cleanly in the 1D shock-free tests. But the tests supply the characteristic trajectories from a Lagrangian solver, and the paper never shows how an Eulerian-only user obtains x^n for the DMD algorithm.\n\nWhat is actually new: applying DMD, rather than POD, in the Lagrangian frame, with the characteristic coordinate as an explicit observable, and a level-set variant for 2D linear advection. The Koopman framing is standard but appropriate. The numerical experiments are honest and well presented: linear advection, advection-diffusion, inviscid and viscous Burgers, all trained on the first 250 snapshots and evaluated in the extrapolating regime. Lagrangian DMD and Lagrangian POD both capture the dynamics, while Eulerian DMD/POD fail. The paper explicitly acknowledges the shock-free restriction and states that shocks are left to future work. The computational table supports the iteration-free efficiency claim.\n\nThe main soft spot is load-bearing: the method needs accurate characteristic trajectories x^n, and the paper does not demonstrate that these can be recovered from Eulerian snapshots. Section 3.1 mentions that an optimal Lagrangian basis can be constructed for POD following [30, Sec. 3.3], but no analogous construction or test is given for Lagrangian DMD. Since DMD is usually marketed as a purely data-driven, equation-free method, this gap matters. A user with standard Eulerian snapshot data has no validated path to the required observables. Secondary soft spots: the abstract claims applicability to general nonlinear advection-diffusion flows, but all tests are 1D and shock-free; no code is shipped; and the error bound in Theorem 4.1 comes from the authors' companion paper [20], so it is not a standalone contribution. None of these are disqualifying, but the Eulerian-only reconstruction issue should be fixed before the method is advertised as ready-to-use.\n\nFor a ROM researcher working on advection-dominated transport, this paper is worth reading and citing as prior art for Lagrangian observables in DMD. It deserves a serious referee: the idea is sensible, the exposition is clear, and the missing piece is a well-defined technical challenge rather than a hidden fatal flaw. I would accept it for peer review and ask for a demonstration of the Eulerian-data route or a clear statement that Lagrangian HFM data are required.","headline":"A genuinely useful incremental extension of Lagrangian POD to DMD, with clean shock-free tests—but the crucial Eulerian-only-data route for getting characteristic trajectories is never demonstrated, so its central promise outruns its evidence.","tokens_in":14807,"tokens_out":2186,"would_cite":true,"duration_ms":28226,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65M99"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that running DMD on the Lagrangian observable $(x,u)$, where $x$ tracks moving characteristic lines, builds reduced-order models that stay accurate beyond the training window for shock-free advection-dominated problems.","keywords":["dynamic mode decomposition","Koopman operator","Lagrangian framework","reduced-order model","advection-dominated problems","proper orthogonal decomposition","characteristic lines","extrapolation"],"falsifier":"Run the same Lagrangian-DMD pipeline on a problem in which characteristic lines cross, such as inviscid Burgers evolved past the shock-formation time, and record the global truncation error $E^n$; the central claim would be falsified if the error stops following the Theorem 4.1 bound and grows without control or produces unphysical oscillations, since the paper's own conclusion states that all its tests are shock-free and that the Lagrangian grid may entangle.","tokens_in":13851,"feed_emoji":"🌊","tokens_out":12771,"duration_ms":119506,"temperature":0.7,"pith_summary":"Conventional dynamic mode decomposition (DMD) and proper orthogonal decomposition (POD) fail for advection-dominated problems because their SVD bases are extracted from the region a traveling wave has already occupied, so the basis cannot represent the wave once it moves on. This paper claims that the fix is to change coordinates: run DMD on the Lagrangian observable $y=(x,u)$, where $x$ tracks the positions of characteristic lines and $u$ is the solution value on those moving lines. In four shock-free numerical tests, linear advection, linear advection-diffusion, and inviscid and viscous Burgers, the Lagrangian DMD captures the solution beyond the 250-snapshot training window ($t>0.25$), where its Eulerian counterparts produce oscillatory and unphysical profiles. The proposed method is iteration-free, so it builds and evaluates the ROM more cheaply than Lagrangian POD while matching or beating its accuracy.","feed_headline":"Lagrangian DMD predicts advection-dominated flows past training data","feed_subtitle":"Conventional DMD and POD fail once the wave travels beyond the training window; Lagrangian DMD stays accurate.","key_machinery":"The load-bearing object is the observable vector $y^n=(x^n,u^n)$ built from the semi-Lagrangian formulation of the advection-diffusion equation, which tracks characteristic lines through the ordinary differential equation $dX/dt=f(u(X(t),t))$. The DMD algorithm approximates the Koopman operator, the infinite-dimensional linear operator that advances observable functions of the state, restricted to the invariant subspace spanned by these observables under Assumption 3.1. The characteristic-line positions are what do the work: they translate the moving wave into a frame in which the low-rank SVD basis does not have to chase the signal, which is exactly the failure mode diagnosed for Eulerian DMD in Section 2.3.","core_discovery":"The paper's central discovery is that the translational difficulty of advection-dominated flows is a coordinate problem, not a fundamental limit of DMD or POD. By choosing the observable $y^n=(x^n,u^n)$, the vector of characteristic-line positions together with solution values on those lines, the authors obtain a finite-dimensional Koopman-invariant subspace in which the moving wave is stationary relative to the basis. The DMD algorithm (Algorithm 3.1) then approximates the Koopman operator restricted to this subspace, and future states are predicted analytically by $y^n=\\Phi \\Lambda^n b$. In the tested shock-free regimes the resulting ROM remains accurate in the extrapolating mode, i.e., for $t>0.25$, and the error bound of Theorem 4.1 gives a computable estimate of how long that accuracy lasts.","pith_inferences":["If characteristic positions can be recovered from Eulerian-only data by tracking wave crests, level sets, or feature velocities, the same ROM construction should work without a dedicated Lagrangian solver; the paper outlines this route in Section 3.1 but does not test it, so this remains an inference about an untested extension.","SVD-based ROMs for moving fronts have usually been repaired with local bases, domain decomposition, or multiresolution filtering; the Lagrangian-observable choice suggests a single global basis can suffice for a wider class of problems, provided the characteristic grid does not entangle.","A natural stress test beyond the four textbook cases is to add small noise or slightly random initial conditions and watch whether the Koopman-invariant-subspace assumption degrades gracefully; the paper does not report such a test, so the practical robustness of the extrapolating mode is still open."],"forward_implications":["For shock-free advection-dominated problems, a single low-rank basis constructed on $(x,u)$ extrapolates accurately far beyond the training window $t\\le 0.25$, where Eulerian DMD and POD fail.","Because future states come from one direct evaluation of the prediction formula, Lagrangian DMD requires no iteration in the low-dimensional space, making it the cheapest ROM among those compared while matching Lagrangian POD's accuracy in most tests.","The error bound in Theorem 4.1 provides an a priori estimate of the observable's error, allowing one to design a hybrid scheme that runs the high-fidelity model for a short time and then switches to the ROM for long-time prediction.","The level-set experiment in the appendix shows that the same Lagrangian-observable idea can be applied to conservation laws by reformulating them as two-dimensional linear transport, capturing the inviscid Burgers solution at rank $r=3$."],"supporting_citations":[{"why":"Supplies the DMD algorithm and the definition of the Koopman operator that the Lagrangian DMD approximates.","marker":"[10]"},{"why":"Introduces Lagrangian POD, the precedent for using characteristic lines as coordinates and the method compared against.","marker":"[30]"},{"why":"Provides the error bound used in Theorem 4.1 to estimate how long the ROM's predictions can be trusted.","marker":"[20]"},{"why":"Supports Assumption 3.1, that observables in a Koopman-invariant subspace admit a finite-dimensional linear representation.","marker":"[31]"},{"why":"Documents the grid-distortion instability of Lagrangian POD in low-dimensional subspaces, which the iteration-free DMD variant avoids.","marker":"[32]"}],"fun_headline_variants":["Lagrangian DMD sees beyond the training horizon","Characteristic lines give DMD predictive power","Shift to Lagrangian: DMD handles advection","Riding characteristics: DMD for moving flows"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a small set of fixed modes can exactly represent the coupled motion of the moving grid positions and the solution values on them, and that the training data reliably supply those moving grid positions; the tests obtain them from a Lagrangian solver, so the method's success depends on that supply.","fun_headline_variants_meta":{"raw":{"variants":["Lagrangian DMD sees beyond the training horizon","Characteristic lines give DMD predictive power","Shift to Lagrangian: DMD handles advection","Riding characteristics: DMD for moving flows"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000271,"raw_usage":{"total_tokens":1581,"prompt_tokens":847,"completion_tokens":734,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":463,"completion_tokens_details":{"reasoning_tokens":676}},"tokens_in":463,"tokens_out":734,"duration_ms":8379,"temperature":1.0,"reasoning_tokens":676,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:06:14.720215+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same Lagrangian-DMD pipeline on a problem in which characteristic lines cross, such as inviscid Burgers evolved past the shock-formation time, and record the global truncation error $E^n$; the central claim would be falsified if the error stops following the Theorem 4.1 bound and grows without control or produces unphysical oscillations, since the paper's own conclusion states that all its tests are shock-free and that the Lagrangian grid may entangle.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the DMD algorithm and the definition of the Koopman operator that the Lagrangian DMD approximates."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces Lagrangian POD, the precedent for using characteristic lines as coordinates and the method compared against."},{"cited_title":"Benner, S","cited_arxiv_id":null,"evidence_quote":"Provides the error bound used in Theorem 4.1 to estimate how long the ROM's predictions can be trusted."},{"cited_title":"Mezi´ c, Spectral properties of dynamical systems, mode l reduction and decompositions, Nonlin","cited_arxiv_id":null,"evidence_quote":"Supports Assumption 3.1, that observables in a Koopman-invariant subspace admit a finite-dimensional linear representation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents the grid-distortion instability of Lagrangian POD in low-dimensional subspaces, which the iteration-free DMD variant avoids."}],"review_version":1}